Balancer

4 deployments · $123.9M aggregate TVL · Dexs

Deployments

Each deployment is rated independently. Pick one to see its rating, risk analysis, and stage.

TVL $85.3M

Type Dexs

Chains Ethereum, Monad, Arbitrum, Base, Hyperliquid L1 +4

View on DeFiLlama ↗

Control Exit Autonomy Open Access Verifiable

Tier bronze AI consensus on at least one dimension
Verifiability tentative Audits exist but audit-to-deploy drift unresolved
Control orange · 2/3 Control unresolved
Ability to exit tentative Exits pausable; recovery mode admin-gated
Autonomy red · 2/3 Rate-provider principal risk
Open Access unknown · 2/3 Frontend access unverified

Control criteria

Upgradeability Unknown Bug bounty immunefi.com Governance forum forum.balancer.fi Docs docs.balancer.fi

About

Balancer V3 is a decentralized automated market maker (AMM) that separates pool logic from token management via its Vault architecture. Users can add or remove liquidity from multi-token pools (up to 8 tokens) using proportional, single-token, or custom operations. The Vault handles all token transfers, balance accounting, decimal scaling, and rate scaling for yield-bearing tokens, while individual pools implement their own swap invariants (weighted, stable, custom).

Risk analysis

One card per dimension, sorted by severity. Only Verifiability and Autonomy carry automated signals in Phase 0. See methodology for scope.

Audit a dimension yourself · DEFI@home Contribute an LLM-run assessment — any model, any dimension. Three agreeing runs merge automatically into the public record.

DEFI@home is a distributed audit network modeled on SETI@home: instead of CPU cycles, it crowdsources LLM reasoning. Paste a slice prompt into Claude, ChatGPT, Gemini, or any browsing-capable model, and submit the JSON output as a pull request. The quorum bot merges it once ≥3 independent runs (from different models) reach the same grade — no single model, and no single contributor, can move the needle alone. How it works →

Address discovery 5 addresses on file · 0 runs Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v3
- protocol.name:              Balancer V3
- protocol.chains:            Ethereum, Monad, Arbitrum, Base, Hyperliquid L1, Optimism, Avalanche, Plasma, xDai
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               [
  {
    "chain": "Ethereum",
    "address": "0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6",
    "role": "Vault Authorizer"
  },
  {
    "chain": "Ethereum",
    "address": "0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f",
    "role": "Balancer DAO Governance (Timelock)"
  },
  {
    "chain": "Ethereum",
    "address": "0x0EF2e7648A9b354822ed0EE9070A2E37332fd242",
    "role": "Emergency SubDAO"
  },
  {
    "chain": "Ethereum",
    "address": "0xA216683066513673FE033917bD5b11f871c65353",
    "role": "Timelock"
  },
  {
    "chain": "Ethereum",
    "address": "0xa731C23D7c95436Baaae9D52782f966E1ed07cc8",
    "role": "Protocol fee controller"
  },
  {
    "chain": "Ethereum",
    "address": "0xba100000625a3754423978a60c9317c58a424e3D",
    "role": "BAL governance token"
  }
]

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
- https://defipunkd.com/address/1/0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6 (Vault Authorizer)
- https://defipunkd.com/address/1/0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f (Balancer DAO Governance (Timelock))
- https://defipunkd.com/address/1/0x0EF2e7648A9b354822ed0EE9070A2E37332fd242 (Emergency SubDAO)
- https://defipunkd.com/address/1/0xA216683066513673FE033917bD5b11f871c65353 (Timelock)
- https://defipunkd.com/address/1/0xa731C23D7c95436Baaae9D52782f966E1ed07cc8 (Protocol fee controller)
- https://defipunkd.com/address/1/0xba100000625a3754423978a60c9317c58a424e3D (BAL governance token)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: DISCOVERY

You are a cataloguer, not a judge. Your job is to surface every contract address that could plausibly belong to this protocol's control or fund-holding surface, each backed by a citation. `grade` is ALWAYS `"unknown"` for discovery submissions — there is no green/orange/red rubric here. The five evaluation slices that run after you (control, ability-to-exit, autonomy, open-access, verifiability) consume your output via the addressBook ratchet — every address you record becomes a pre-built surfacer URL on the next run; every address you miss costs them a tool call.

Width beats depth. A `role: "other"` entry with one cited URL beats omitting it. Downstream slices will discard out-of-scope entries; they cannot rediscover what you fail to enumerate without paying the same cost again.

(Step 0 capability probe lives in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every D-code below must appear in evidence[] OR unknowns[])

- **D1. Block-explorer name-tag search per chain.** For each chain in `protocol.chains`, search the canonical block explorer for the protocol's name tag — `https://etherscan.io/searchHandler?term=<query>` and the per-chain explorers (basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network). When direct fetch is blocked, use `site:<explorer> <protocol_name>` via search grounding. Record every address that surfaces with the protocol's tag, plus neighbouring "Token Contract" / "Multisig" / "Timelock" labels.

- **D2. Official deployments doc.** From `protocol.website` and the docs site, locate the canonical "Deployed contracts" / "Addresses" / "Contracts" / "Deployments" page (often at `/docs/deployments`, `/docs/addresses`, `/dashboard/contracts`). Cite the URL, record every address listed with its named role, and set `protocol_metadata.deployed_contracts_doc` to this URL.

- **D3. Audit PDFs.** From `protocol.audit_links` (and any audits surfaced by D2), open each. Most reports include a "Scope" / "Contracts in scope" address table in the first 5 pages. Extract every in-scope address with its labelled role. If the audit predates the current deployment, record the addresses anyway with role suffixed `(audit-era)` so downstream slices know to re-verify.

- **D4. GitHub deployment artifacts.** From `protocol.github`, walk the repo at a pinned commit SHA looking for: Foundry `broadcast/<script>/<chainId>/run-latest.json` (`transactions[].contractAddress` per chainId); hardhat-deploy `deployments/<network>/<Contract>.json` (`address` field); manual indexes (`deployments.json`, `addresses.json`, `contracts.json`, `networks.json`); markdown indexes (`docs/deployments.md`, `README.md` tables). Cite the file URL with the commit SHA; pin SHAs (`?ref=<sha>`) so the citation is content-addressed.

- **D5. Multi-chain enumeration.** If `protocol.chains.length > 1`, repeat D1–D4 per chain. Cross-chain deployments of the same logical contract get SEPARATE `admin_addresses[]` entries — one per chain. The chain field is part of the identity; do not collapse. If a chain has zero results, record `"D5: chain <name>: zero addresses surfaced from <sources tried>"` in unknowns[].

- **D6. Factory-discovered children.** For factory addresses surfaced in D1–D4, fetch the enumeration view via the read API (`/api/contract/read?...&method=allPools` / `getPool` / `getMarket` / `getVault`) and record each child with role like `"pool (from factory <0xFactory>)"`. **Cap at 50 children per factory.** Protocols with thousands of pools (Uniswap, Sushi) need dedicated ingestion — record the factory + the cap notice in unknowns[].

- **D7. Role taxonomy.** Every `admin_addresses[]` entry's `role` uses this controlled vocabulary (free-text suffixes OK for disambiguation, e.g. `"multisig (treasury)"`, but the leading token must match):

    `owner | admin | proxy_admin | governor | timelock | guardian | multisig | treasury | oracle | factory | router | token | pool | vault | other`

  Tentative classifications are encouraged. `actor_class` ∈ `eoa | multisig | timelock | governance | unknown` — use `unknown` when you found the address but didn't read its bytecode.

- **D8. Ratchet output integrity.** Every address in `admin_addresses[]` must trace to ≥1 fetched URL in evidence[]. Snippet-only sightings go in unknowns[] with a `D8` code, NOT in admin_addresses[].

### Discovery rationale framing

- `rationale.findings`: one entry per D-code, terse, factual. Per-address detail belongs in evidence[] and admin_addresses[], not here. Example: `"D1: 8 addresses surfaced from etherscan.io name-tag search for 'Aave V3'"`.
- `rationale.steelman`: ALWAYS null.
- `rationale.verdict`: one short line summarizing what corpora were walked and how many addresses were catalogued.
- `headline`: factual and quantitative — `"24 contracts catalogued across Ethereum, Arbitrum, and Base; 6 governance/admin and 18 protocol contracts."`.
- `short_headline`: under 60 chars — `"24 contracts across 3 chains"`.

### What discovery is NOT

- Not a verdict slice. `grade` must be `"unknown"`.
- Not exhaustive enumeration of leaf assets — record the factory + cap and move on (see D6).
- Not classification of trust assumptions — whether a multisig threshold is safe / timelock delay is sufficient / proxy admin is an EOA is the control slice's job.
- Not address-book reconciliation: when addressBook is non-empty, EXTEND it (find addresses prior runs missed) rather than re-cite the same addresses; re-cite only when you have new evidence for a refined role.

### protocol_metadata side-effects

While walking the corpora, populate every `protocol_metadata` field you can support with citations: `github`, `docs_url`, `audits` (one per D3 audit walked), `governance_forum`, `bug_bounty_url`, `security_contact`, `deployed_contracts_doc` (URL from D2), `upgradeability` (best-effort), and `about` (2–4 sentences sourced from docs/website, not memory). Discovery is the natural home for these — evaluation slices should not have to rediscover them.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "discovery",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Verifiability ✓ 3/3 models agree AI-only weak orange — weak consensus margin; only 0/3 sources have a public chat share link; total support weight 0.31 below confidence floor (1.5) Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v3
- protocol.name:              Balancer V3
- protocol.chains:            Ethereum, Monad, Arbitrum, Base, Hyperliquid L1, Optimism, Avalanche, Plasma, xDai
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               [
  {
    "chain": "Ethereum",
    "address": "0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6",
    "role": "Vault Authorizer"
  },
  {
    "chain": "Ethereum",
    "address": "0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f",
    "role": "Balancer DAO Governance (Timelock)"
  },
  {
    "chain": "Ethereum",
    "address": "0x0EF2e7648A9b354822ed0EE9070A2E37332fd242",
    "role": "Emergency SubDAO"
  },
  {
    "chain": "Ethereum",
    "address": "0xA216683066513673FE033917bD5b11f871c65353",
    "role": "Timelock"
  },
  {
    "chain": "Ethereum",
    "address": "0xa731C23D7c95436Baaae9D52782f966E1ed07cc8",
    "role": "Protocol fee controller"
  },
  {
    "chain": "Ethereum",
    "address": "0xba100000625a3754423978a60c9317c58a424e3D",
    "role": "BAL governance token"
  }
]

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
- https://defipunkd.com/address/1/0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6 (Vault Authorizer)
- https://defipunkd.com/address/1/0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f (Balancer DAO Governance (Timelock))
- https://defipunkd.com/address/1/0x0EF2e7648A9b354822ed0EE9070A2E37332fd242 (Emergency SubDAO)
- https://defipunkd.com/address/1/0xA216683066513673FE033917bD5b11f871c65353 (Timelock)
- https://defipunkd.com/address/1/0xa731C23D7c95436Baaae9D52782f966E1ed07cc8 (Protocol fee controller)
- https://defipunkd.com/address/1/0xba100000625a3754423978a60c9317c58a424e3D (BAL governance token)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: VERIFIABILITY

Evaluate whether an outsider can independently confirm what the deployed code does.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- V1. For each address you assess: is the bytecode verified on the chain's block explorer? Record the "Contract Source Code Verified" indicator. https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x... returns this as a top-level "verified" boolean plus "abiSource" ("etherscan" / "sourcify") and an inline ABI — useful when the explorer page is rate-limited. If the contract is a proxy, verify BOTH the proxy contract AND the current implementation contract. The same /api/contract/abi response auto-resolves proxies and includes a "proxy.implementation" address when present, so one call covers V1 + V6 in one shot. An explorer "Similar Match" on a well-known proxy pattern (Aragon AppProxyUpgradeable, ERC1967Proxy, OssifiableProxy, OZ TransparentUpgradeableProxy) is expected for that pattern and does NOT count as a verification gap on its own — what matters is that the implementation is independently verified.
- V2. Source-to-repo correspondence: for each verified contract, attempt to find a matching commit in the linked GitHub repos. Record evidence[].commit on a match. Independent compile/bytecode-match is NOT required for green — a recognized public repo whose structure and file contents correspond to the explorer-visible source is sufficient. If you did not pin a commit SHA or run a bytecode diff, record that plainly in unknowns[] and proceed; it is a scope limit, not a downgrade signal.
- V3. Audit coverage: for each URL in protocol.audit_links, open it and record: auditor name, audit date, the specific contracts / commit in scope. Flag audits that predate the current deployment by >6 months without a follow-up review.
- V4. Auditor recognition: the following firms are broadly recognized in Solidity: Trail of Bits, Zellic, Spearbit, OpenZeppelin, ConsenSys Diligence, Certora (formal verification), Quantstamp, Halborn, Peckshield, Sigma Prime, ChainSecurity, Ackee Blockchain, MixBytes, Statemind. Unknown firms are orange-at-best for any green-grade claim. Name the firm explicitly in evidence[].
- V5. Post-audit drift: compare the most recent in-scope audit(s) against the currently-deployed source, weighted by what each contract does and by what the changes actually contain. SCOPE — drift only downgrades the grade when ALL of the following hold: (i) the drifting contract is fund-custody / settlement / accounting-critical (NOT a peripheral router, lens, quoter, or pure-view contract that holds no balances); (ii) the changes are material — new functions, modified access control, modified accounting, modified fund flow — and not refactors / struct relocations / import reorgs / build or CI fixes / formatting; (iii) no later audit, fix-audit, or differential audit from a recognized firm covers the changed files (audits often pin a pre-fix commit while a follow-up reviews the delta — match by file scope, not by commit-hash equality). When you cite drift as a downgrade reason, name the specific behavior change (function added, role granted, accounting formula altered) — "N commits ahead" or "+X/-Y LOC" alone is not evidence of material drift. If you have not sampled the diff content (e.g. via the GitHub compare view or the top commits in the window), record drift as an unknowns[] entry rather than auto-downgrading; commit-count and LOC are starting signals, not findings.
- V6. Implementation vs proxy: a verified proxy with an unverified implementation is effectively unverified. State whether the implementation is verified separately.

EVIDENCE DISCIPLINE (read before writing findings[]):
- Do not assert a specific deploy-commit SHA, bytecode equivalence, or "identical to audited commit" unless you actually fetched the artifact that shows it (e.g., a deployed-addresses JSON you opened, an explorer page you read). Inferred or plausible matches belong in unknowns[], never in findings[] or evidence[].
- Evidence[] entries must correspond to pages/files you actually retrieved this run. A URL you did not open is not evidence.

Then write the steel-man section per Hard Rule 11.

Grade rules:
- green   = deployed bytecode verified on the explorer (proxy AND implementation if proxied; "Similar Match" on a standard proxy pattern is fine per V1), a public source repo exists whose contents correspond to the explorer-visible source, AND ≥1 audit from a recognized firm covering the currently-deployed contracts (≤6 months of drift OR drift was re-audited). A missing local compile-match is not a downgrade — record it in unknowns[] and still grade green if the other conditions hold.
- orange  = verified but with visible drift from the public repo, OR audit scope is stale relative to deployment, OR only minor / unknown-firm audits exist, OR only some of the main contracts are verified, OR proxy verified but implementation only partially verified.
- red     = unverified bytecode (or verified proxy with unverified implementation), OR no audit in protocol.audit_links, OR no public repo.
- unknown = reserved for when the protocol's verifiability posture genuinely cannot be assessed (e.g., explorer and repo both inaccessible for this protocol). Do NOT use unknown merely because you, the analyst, could not run a particular check such as a bytecode diff — that goes in unknowns[] while the grade is still assigned from the evidence you do have.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "verifiability",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Control 2/3 submitted Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v3
- protocol.name:              Balancer V3
- protocol.chains:            Ethereum, Monad, Arbitrum, Base, Hyperliquid L1, Optimism, Avalanche, Plasma, xDai
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               [
  {
    "chain": "Ethereum",
    "address": "0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6",
    "role": "Vault Authorizer"
  },
  {
    "chain": "Ethereum",
    "address": "0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f",
    "role": "Balancer DAO Governance (Timelock)"
  },
  {
    "chain": "Ethereum",
    "address": "0x0EF2e7648A9b354822ed0EE9070A2E37332fd242",
    "role": "Emergency SubDAO"
  },
  {
    "chain": "Ethereum",
    "address": "0xA216683066513673FE033917bD5b11f871c65353",
    "role": "Timelock"
  },
  {
    "chain": "Ethereum",
    "address": "0xa731C23D7c95436Baaae9D52782f966E1ed07cc8",
    "role": "Protocol fee controller"
  },
  {
    "chain": "Ethereum",
    "address": "0xba100000625a3754423978a60c9317c58a424e3D",
    "role": "BAL governance token"
  }
]

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
- https://defipunkd.com/address/1/0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6 (Vault Authorizer)
- https://defipunkd.com/address/1/0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f (Balancer DAO Governance (Timelock))
- https://defipunkd.com/address/1/0x0EF2e7648A9b354822ed0EE9070A2E37332fd242 (Emergency SubDAO)
- https://defipunkd.com/address/1/0xA216683066513673FE033917bD5b11f871c65353 (Timelock)
- https://defipunkd.com/address/1/0xa731C23D7c95436Baaae9D52782f966E1ed07cc8 (Protocol fee controller)
- https://defipunkd.com/address/1/0xba100000625a3754423978a60c9317c58a424e3D (BAL governance token)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: CONTROL

Evaluate who can change the protocol's rules, how fast, and how broadly.

(Step 0 capability probe and the off-chain-only fallback live in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[])

- **C1.** For each address you assess: who is the contract owner / admin / pendingAdmin / governor — read these via the block explorer's "Read Contract" tab OR `https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=owner` (BARE method names: `&method=owner`, `&method=admin`, `&method=pendingOwner`, `&method=governor`). For Safes use `/api/safe/owners?chainId=<id>&address=0x...`. When a protocol has multiple major versions deployed (v2/v3/v4), perform C1 reads on the NEWEST deployment separately — newer deployments often have weaker control surfaces than the legacy core.
- **C2.** Upgrade mechanism: transparent proxy / UUPS / Beacon / Diamond / immutable. Identify the proxy admin address. Check upgradeability of GOVERNANCE contracts too — a Governor / Aragon Voting / OZ Governor is often itself a proxy whose admin is the Timelock. Asymmetry: when fund-holding cores are immutable AND governance has no admin path that reaches them, an upgradable Governor/Timelock is T3-only and must NOT drag the verdict below green on that basis alone (see grade rules and the "immutable cores" caveat). Only call upgradability "mixed" if you can name a concrete function on the upgradable surface that reaches T1 or T2 on user funds.
- **C3.** EXECUTION PATH (enumerate every stage, in order, with delays in seconds). The operative path is usually a chain (voting → scheduler → timelock → executor; or governor → queue → execute; or Aragon Voting → DualGovernance → EmergencyProtectedTimelock → AdminExecutor). For each stage, record (a) the contract address, (b) the delay constant name + value in seconds, (c) the URL you read it from (block-explorer Read Contract OR `/api/contract/read?...&method=MIN_DELAY&block=<n>`). Do NOT stop at the first timelock-shaped contract — if its admin is itself called by another contract, keep walking. The grading delay is the SUM OF DELAYS ON THE UNCONTESTED FAST PATH (shortest time a proposal with no opposition can go from submission to executable). Dynamic / contested extensions (veto signaling, rage quit, escrow delay) are modifiers, not the basis — note them separately.
- **C4.** Enumerate EVERY multisig with reachable control — main proxy admin, emergency activation, emergency execution, reseal / pause, gate-seal committees, tiebreaker, per-module admins. For each Safe, fetch threshold + owners + version via `/api/safe/owners?chainId=<id>&address=0x...` (response includes raw eth_call data, so the URL is citable evidence). Enumerate ops/council/incentives multisigs even when off the upgrade path — record their scope so a reader can see they are NOT on the upgrade path. For each: (a) address, (b) threshold / total signers, (c) signer identities classified as insider (team, paid auditors under ongoing engagement, mandated service providers) vs non-insider (independent community members, unaffiliated researchers), (d) the specific power held (upgrade, pause, parameter, etc.).
- **C5.** On-chain governance: Governor / GovernorBravo / OZ Governor / Aragon Voting with token-weighted voting? Record proposal threshold, voting period, quorum, and the timelock delay between queue and execute. Every numeric constant must come from a Read Contract call you can link to, or be in unknowns[] with the C-code. If votingDelay / votingPeriod are denominated in BLOCKS, convert to seconds at the chain's CURRENT block time (Ethereum mainnet ≈ 12s post-Merge, not the 15s in older Compound/Bravo deployments) — cite both block count and converted seconds.
- **C6.** EMERGENCY POWERS: separate emergency-pause / guardian role with a different time cap or different actor than the main upgrade authority? Record it explicitly.
- **C7.** POWER TIER (blast radius). For each privileged path in C3–C6, classify the WORST thing that path can do, choosing the highest applicable tier. Cite the specific function name and any on-chain bound — tier claims without a named function are unsupported.
    - **T1 — FUND-CRITICAL**: replace implementation of contracts holding user funds; change AMM math / accounting / collateral logic; mint unbacked debt or shares; pause withdrawals; drain user-fund treasury; change oracle to attacker-controlled source; replace upgrade admin with EOA.
    - **T2 — ECONOMICALLY MATERIAL**: change fee parameters within bounded ranges; redirect protocol fees; add/remove markets / collateral types; bounded inflation or token mint within hard-capped schedule; spend protocol-owned (non-user) treasury.
    - **T3 — GOVERNANCE-INTERNAL**: change voting rules, quorum, voting period, proposal threshold; upgrade the Governor itself; rotate Timelock admin; mint governance tokens within a capped annual schedule.
    - **T4 — OPERATIONAL**: incentives distribution, grants, ENS / frontend canonicalization, deployment coordination, periphery router deprecation.
  The grade is set by the HIGHEST tier reachable on the uncontested fast path, not the median. State the tier and the binding function in the verdict.

### Read Contract discipline (applies to C3, C4, C5)

Every numeric constant cited (timelock delays, voting periods, multisig thresholds, quorum percentages) must come from EITHER (a) a block-explorer Read Contract URL, OR (b) a DeFiPunkd `/api/contract/read` or `/api/safe/owners` URL (preferred with `&block=<n>` for content-addressed evidence), OR (c) an unknowns[] entry with the C-code. Docs / blog posts are corroboration only — they cannot be the sole citation for a value that is also readable on-chain.

### Off-chain-only substitute hierarchy (when grading_basis="off-chain-only" — see preamble Rule 16)

When on-chain reads were genuinely unreachable this run, eligible off-chain substitutes in priority order:
1. Linked audit PDFs (admin roles, multisig members, timelock delays usually enumerated).
2. Governance forum posts that quote constants from a successful on-chain proposal (cite post URL + linked execution-tx URL).
3. Official protocol docs pages with named addresses and roles (must be on a domain owned by the protocol).
4. GitHub README / SECURITY.md / governance/*.md at a pinned commit SHA.

Forbidden substitutes: third-party blog posts, X / Twitter threads, search-result snippets, model memory. Required degradation: any C-code citing a numeric constant from docs/forum/audit prose ONLY must also carry an `unknowns[]` entry with `-offchain` suffix noting "value not re-read on-chain in this run; corroboration only".

### Grade rules (apply the timelock bar conditional on the highest C7 tier reachable on the fast path)

Security Council standard (used below): a multisig qualifies as "Security Council" only if ALL of: ≥7 signers, ≥51% threshold, ≥50% non-insider signers, every signer publicly announced. Failing any criterion = NOT a Security Council, regardless of signer reputation.

- **green**: highest reachable tier is T3 or T4 regardless of timelock; OR T2 reachable with uncontested-fast-path delay ≥7 days; OR T1 reachable only via immutable contracts (T1 is unreachable); OR T1 reachable with uncontested-fast-path delay ≥7 days combined with a Security Council multisig; OR T1 reachable with uncontested-fast-path delay ≥7 days through active on-chain governance with broad token distribution.
- **orange**: T2 reachable with uncontested-fast-path delay >0 but <7 days; OR T1 reachable with uncontested-fast-path delay >0 but <7 days; OR a multisig failing one or more Security Council criteria sits on a T1/T2 path; OR unclear upgrade authority on a T1/T2 path; OR governance with very short timelock or low quorum on a T1/T2 path.
- **red**: T1 reachable with no timelock by a single EOA or 2-of-3 multisig; OR a T1 upgrade admin that is not a smart contract you can audit.
- **unknown**: completed the checklist but still cannot determine the upgrade authority OR cannot classify the highest tier reachable on the main contracts.

Tiering caveats:
- "Bounded" must be enforced ON-CHAIN to count as T2. A function that sets fees with no upper-bound check is T1 — cite the bound check.
- Recurring T2 economic extraction (e.g. fee redirect with no rate limit) approaches T1 over time. A single proposal that can permanently redirect all future revenue is T1.
- T3 assumes the governance contract cannot itself authorize a T1/T2 action without going through the same timelock. If governance can self-upgrade to bypass the timelock, T3 collapses into T1.
- Do not downgrade tier by hand-waving ("realistically governance would never…"). Tier on what the contract permits, not what feels likely.

Notes:
- **Dynamic / dual-governance timelocks** (Lido, Compound escrow veto): the rubric grades on the uncontested path because that is the path most upgrades take. A dynamic extension that fires only under stake-weighted opposition is a real protection — name it in the green steel-man, but it does not lift an orange fast path into green; state the tension in the verdict.
- **Immutable cores with upgradable governance** (Uniswap-style): if fund-holding contracts are immutable and have no admin-reachable function moving / freezing / re-routing user funds, the highest reachable tier on the upgrade path is T3 — green regardless of timelock. Don't grade this orange just because the Governor is a proxy — that's a C2 fact, not a downgrade. Downgrade only applies if you can cite a concrete function on the upgradable surface that reaches T1 or T2 (privileged hook, upgradable factory controlling fund-routing, fee-switch redirecting protocol revenue without bound).
- **The 7-day bar** reflects the exit-window standard — users need notice after a queued upgrade to withdraw if they disagree. The ability-to-exit slice grades the exit side; this slice grades the delay side; both must hold for users to actually benefit from the delay.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "control",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Ability to exit ✓ 3/3 models agree AI-only weak orange — weak consensus margin; only 0/3 sources have a public chat share link; total support weight 0.09 below confidence floor (1.5) Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v3
- protocol.name:              Balancer V3
- protocol.chains:            Ethereum, Monad, Arbitrum, Base, Hyperliquid L1, Optimism, Avalanche, Plasma, xDai
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               [
  {
    "chain": "Ethereum",
    "address": "0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6",
    "role": "Vault Authorizer"
  },
  {
    "chain": "Ethereum",
    "address": "0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f",
    "role": "Balancer DAO Governance (Timelock)"
  },
  {
    "chain": "Ethereum",
    "address": "0x0EF2e7648A9b354822ed0EE9070A2E37332fd242",
    "role": "Emergency SubDAO"
  },
  {
    "chain": "Ethereum",
    "address": "0xA216683066513673FE033917bD5b11f871c65353",
    "role": "Timelock"
  },
  {
    "chain": "Ethereum",
    "address": "0xa731C23D7c95436Baaae9D52782f966E1ed07cc8",
    "role": "Protocol fee controller"
  },
  {
    "chain": "Ethereum",
    "address": "0xba100000625a3754423978a60c9317c58a424e3D",
    "role": "BAL governance token"
  }
]

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
- https://defipunkd.com/address/1/0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6 (Vault Authorizer)
- https://defipunkd.com/address/1/0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f (Balancer DAO Governance (Timelock))
- https://defipunkd.com/address/1/0x0EF2e7648A9b354822ed0EE9070A2E37332fd242 (Emergency SubDAO)
- https://defipunkd.com/address/1/0xA216683066513673FE033917bD5b11f871c65353 (Timelock)
- https://defipunkd.com/address/1/0xa731C23D7c95436Baaae9D52782f966E1ed07cc8 (Protocol fee controller)
- https://defipunkd.com/address/1/0xba100000625a3754423978a60c9317c58a424e3D (BAL governance token)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: ABILITY-TO-EXIT

Evaluate whether users can withdraw their funds on their own terms, even under adversarial admin conditions.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- E1. Enumerate every user-facing exit function in the main contracts: withdraw, redeem, burn, requestWithdrawal, claim, exit, etc. List them by name. Do NOT treat the contract as a monolith.
- E2. For EACH exit function in E1: identify its access modifiers and any pause guards (e.g. _checkResumed, whenNotPaused, onlyRole). Functions that gate REQUEST PLACEMENT often differ from functions that CLAIM FINALIZED FUNDS — check both separately.
- E3. For each pause guard: identify the role holder (which address holds PAUSE_ROLE / GUARDIAN / etc.) and the maximum pause duration. Specifically check whether PAUSE_INFINITELY (or equivalent uncapped pause) is callable, and which actor can call it (single multisig vs governance vote). For role-holder reads use https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=hasRole&args=0x...,0x... or &method=getRoleAdmin&args=0x.... For "is currently paused" checks use &method=paused or &method=isPaused&args=<resume-code>. Use the BARE method name (no parens). Cite the URL with &block=<n> in evidence[].
- E4. EMERGENCY vs GOVERNANCE pause distinction: many protocols have a fast-acting emergency pause capped at N days and a slow governance pause that can be indefinite. Record both paths separately if present, with their time caps and actor classes.
- E5. Queued redemption: documented maximum queue duration, daily withdrawal caps, whether the queue itself is pausable.
- E6. Forced-exit / escape-hatch / permissionless emergency-exit mechanism for adversarial-admin scenarios.
- E7. Frontend dependency: confirm exit functions are directly callable on-chain (e.g. via Etherscan write tab or a generic wallet) without the project's frontend.

Then write the steel-man section per Hard Rule 11. Common red-vs-orange tension on this slice: indefinite pause exists (suggests red) BUT the realistic emergency path is time-capped AND claims of already-finalized exits are not pause-gated (suggests orange). Resolve this by stating who can do what for how long, not by stopping at the worst-case sentence.

Grade rules:
- green   = permissionless exit; pause is either absent, narrowly scoped to clearly-described emergencies with auto-expiry, or capped at ≤7 days; no frontend dependency for exit; claims of already-finalized exits are not pause-gated under any path.
- orange  = pausable with broad scope OR indefinite pause is reachable only through governance vote (not unilateral admin action), OR queued redemption with documented max > 7 days, OR claims-of-finalized are exempt but new-request placement can be paused indefinitely by governance.
- red     = exit requires admin signature, OR ANY actor (including governance) can pause CLAIMS of finalized exits indefinitely, OR there is no on-chain exit function at all (purely custodial), OR pause is held by a single EOA / 2-of-3 multisig with no time cap.
- unknown = checklist incomplete after checking the sources above.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "ability-to-exit",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Autonomy 2/3 submitted Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v3
- protocol.name:              Balancer V3
- protocol.chains:            Ethereum, Monad, Arbitrum, Base, Hyperliquid L1, Optimism, Avalanche, Plasma, xDai
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               [
  {
    "chain": "Ethereum",
    "address": "0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6",
    "role": "Vault Authorizer"
  },
  {
    "chain": "Ethereum",
    "address": "0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f",
    "role": "Balancer DAO Governance (Timelock)"
  },
  {
    "chain": "Ethereum",
    "address": "0x0EF2e7648A9b354822ed0EE9070A2E37332fd242",
    "role": "Emergency SubDAO"
  },
  {
    "chain": "Ethereum",
    "address": "0xA216683066513673FE033917bD5b11f871c65353",
    "role": "Timelock"
  },
  {
    "chain": "Ethereum",
    "address": "0xa731C23D7c95436Baaae9D52782f966E1ed07cc8",
    "role": "Protocol fee controller"
  },
  {
    "chain": "Ethereum",
    "address": "0xba100000625a3754423978a60c9317c58a424e3D",
    "role": "BAL governance token"
  }
]

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
- https://defipunkd.com/address/1/0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6 (Vault Authorizer)
- https://defipunkd.com/address/1/0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f (Balancer DAO Governance (Timelock))
- https://defipunkd.com/address/1/0x0EF2e7648A9b354822ed0EE9070A2E37332fd242 (Emergency SubDAO)
- https://defipunkd.com/address/1/0xA216683066513673FE033917bD5b11f871c65353 (Timelock)
- https://defipunkd.com/address/1/0xa731C23D7c95436Baaae9D52782f966E1ed07cc8 (Protocol fee controller)
- https://defipunkd.com/address/1/0xba100000625a3754423978a60c9317c58a424e3D (BAL governance token)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: AUTONOMY

Evaluate this protocol's autonomy: can a failure of anything outside its own contracts cause theft or loss of user principal, loss of unclaimed yield, or materially change the protocol's expected performance? "Autonomy" is not "has zero external touchpoints" — it is "external failures are either survivable or recoverable without user loss". This slice adapts DefiScan v1's Autonomy dimension; dependencies are one of several criteria, not the whole frame.

GUARDRAILS (read before grading):
- Category alone (Liquid Staking, Bridge, RWA Lending, Restaking, …) does NOT force a grade. A category is a hint about where to look; the grade must come from the concrete A1–A9 findings below.
- Base-chain consensus (Ethereum PoS, the chain's validator set, the canonical Deposit Contract at 0x00000000219ab540356cBB839Cbe05303d7705Fa) is the SUBSTRATE, not a dependency, for any protocol deployed on that chain. Do not list "depends on Ethereum" as a finding.
- Oracles or other integrations used by DOWNSTREAM protocols that happen to read this protocol's token (e.g. Chainlink stETH/USD consumed by Aave) are NOT this protocol's dependencies. Count only what THIS protocol's contracts call or trust on-chain. If a feed or contract appears in the protocol's docs only as reference material for third-party integrators, EXCLUDE IT ENTIRELY — do not log it as a finding even with a "peripheral" or "referenced only" caveat.
- "Upgradeable admin can change things" belongs to the CONTROL slice; only count it here when the upgrade surface lets governance silently swap an external dependency (see A9).
- Underlying-asset risk in opt-in, isolated markets is NOT autonomy-red on its own. When a protocol wraps third-party yield-bearing assets (LSTs, LRTs, ERC-4626 vaults, lending receipts, restaked tokens) into per-market silos that users explicitly choose, a failure of one underlying does NOT propagate to other markets and is risk the user opted into per-market. Record it under A4 with depth and propagation scope, but do not let it alone drive a red verdict — red requires an external dependency that cross-cuts the protocol or that the user did not opt into at deposit time. A failure mode that is "if the LRT you deposited is hacked, your principal in that LRT-backed market is impaired" is the underlying's autonomy story, not this protocol's; grade it on whether THIS protocol introduces additional dependencies on top.
- Sub-module enumeration is mandatory before grading. If the protocol ships distinct product lines or modules (e.g., a v2 core AMM plus a newer perps/funding-rate module, a lending pool plus a separate vault layer, an L1 core plus a cross-chain extension), enumerate each in findings and grade against the WORST module weighted by its share of TVS or its blast radius. A green core does not rescue an orange/red sub-module; conversely, a small red sub-module with capped TVS may bound the overall grade to orange. Name each module by its on-chain factory or router address. If you do not know whether a module exists, that is an unknowns[] entry, not silence. EXPLICIT TVS WEIGHTING: in the verdict, state each module's approximate share of total protocol TVS (use "~X%" if exact figures unavailable; check DeFiLlama or block-explorer balances on the module's main contract) and how that share informs the weighted grade. Format: "Module A holds ~X% of TVS (grade: <g>); Module B holds ~Y% (grade: <g>); weighted overall = <grade> because <reason>." If a red sub-module holds <5% of total TVS and is capped, the overall grade may be orange; if it holds >25%, the overall grade is red. Do not let qualitative reasoning substitute for the percentages — write the numbers.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. External contract calls. Enumerate every external contract the core contracts call or read from (oracles, price feeds, AMM pools, lending pools, staking/deposit contracts, yield wrappers). For each, identify the address, the provider, and what user-facing function of this protocol would break or mis-price if that external contract paused, mis-reported, or behaved adversarially. Grep for "oracle", "aggregator", "getPrice", "latestAnswer", "chainlink", "pyth", "redstone" as a starting point. To verify an oracle is live and what it currently reports, hit https://defipunkd.com/api/contract/read?chainId=<id>&address=<oracle>&method=latestAnswer (or &method=latestRoundData, &method=getPrice, &method=description; use the BARE method name without parens) and cite the URL — the response includes blockNumber/blockHash and rawReturnData, which is stronger evidence than a docs page about which feed "should" be used.
- A2. Off-chain actor committees reporting INTO the protocol. Oracle committees, guardian multisigs, DAO-selected validator sets acting as protocol reporters, exit-bus signers, fraud-proof challengers. For each, record committee size, quorum, who picks the members, and what mis-reporting could do (mint, burn, finalize withdrawal, freeze). Distinguish this from governance admins (control slice). NOTE ON STAKING PROTOCOLS: validator slashing and node-operator misbehavior are properties of the base-chain substrate, NOT external dependencies, when the operator set is diversified enough that a coordinated failure caps at <5% principal loss. Count validator/operator risk under A2 ONLY if the operator set is small, non-diversified, or lacks bonding / slashing-insurance / diversification mitigations. A curated set of 30+ independent operators with documented diversification falls under the mitigated path; a 3-operator LST does not. Do not cite the protocol's own risk disclosure as evidence that operator failure = principal loss unless you also check the diversification and bond mitigations.
- A3. Bridge / cross-chain messaging dependencies. Only count bridges that carry material TVL or are required for a core user flow. For each, name the bridge operator (canonical L1↔L2, LayerZero, Wormhole, Axelar, custom multisig), the trust model (canonical, optimistic, light-client, guardian set), and what fraction of TVL or users ride on it. "wstETH exists on 15 chains" is not a finding unless material TVL sits there. Before listing any non-primary chain deployment as a dependency, verify it is still operational as of analysis_date — retired or sunset deployments (e.g., Lido-on-Terra, Lido-on-Solana) belong in unknowns[] or should be omitted, never cited as a current dependency.
- A4. Nested collateral / restaking chains. For restaking / LRT / receipt-of-receipt designs: record the depth of the collateral chain, every actor with slashing or freezing power at each level, and whether a failure N levels deep propagates to user principal here.
- A5. Fork lineage (silent check). If DeFiLlama's forkedFrom is non-empty, record it as one finding and move on. If empty, do NOT add a placeholder finding; it adds noise.
- A6. Fallback mechanisms and circuit breakers. What catches an external failure? Sanity-check contracts on oracle reports, rebase bounds, pause paths triggered by bad prices, second-opinion oracles, max-per-block throughput caps, withdrawal queues that absorb bad reports. Record which A1–A4 risks are mitigated by which fallback, and which are unmitigated. For EACH fallback you cite, state its activation status explicitly: (i) LIVE and enforcing on-chain today, (ii) DEPLOYED but not yet wired / activated (e.g., interface exists but the address is zero or the role is unassigned), or (iii) DOCUMENTED / PROPOSED only (forum post, LIP draft, audit pending). Only (i) counts as mitigation for the grade. A fallback in state (ii) or (iii) should be noted but must not reduce the risk in your steel-man or verdict. If you cannot determine activation status, add an unknowns[] entry rather than assume it is live.
- A7. Sequencer / L1-liveness dependency BEYOND the base-chain substrate. SCOPE — sequencer risk only counts here when the protocol IS its own L2/L3 appchain or app-rollup, where the sequencer is part of the protocol's own stack and a freeze is a protocol-level outage. A protocol permissionlessly deployed on a third-party L2 inherits that L2's sequencer as substrate, not an A7 dependency. Record the sequencer/DA trust model when A7 applies.
- A8. Keeper / relayer / off-chain bot liveness. Protocols that need permissionless-but-necessary off-chain actors (liquidation bots, auto-compounders, deposit relayers, intent solvers). Record whether the role is permissionless, what degrades if nobody runs it (yield paused, bad debt accumulates, positions go stale), and whether the failure mode is graceful or catastrophic.
- A9. Governance-mutable dependency surface. Can an admin or DAO action silently INTRODUCE a NEW EXTERNAL dependency — swap the oracle address to a different provider, register a new staking module that calls an untrusted contract, add a new bridge, route SY through a new external vault — without an exit window for users? Check the upgrade / router / module-registry contracts. Answer: which EXTERNAL dependencies are governance-mutable, who holds that power, and whether there is a timelock or exit window. SCOPE LIMIT — read carefully: A9 is about the *external dependency surface*, not the upgrade surface in general. "The proxyAdmin / EOA can upgrade the router implementation to arbitrary bytecode" is a CONTROL-slice finding (admin can rug), NOT an autonomy-A9 finding. A9 fires only when the upgrade specifically swaps out or adds a contract that THIS protocol calls or trusts (e.g., changing the oracle address from Chainlink to a malicious feed, registering a new SY adapter that points to a third-party vault, redirecting a bridge endpoint). If the only finding is "admin can change implementation," do not log it under A9 and do not let it drive the autonomy grade — note it under control instead. The autonomy-relevant version of the same upgrade key is "admin can swap [specific external address X] without timelock"; that requires identifying the specific external dependency that becomes mutable.

STEEL-MAN (per Hard Rule 13): write one-sentence strongest arguments for red, orange, and green using the A1–A9 findings.

IMPACTED TVS ESTIMATE: the headline MUST include a rough impacted-TVS figure — the fraction of protocol TVS that could be lost or frozen if the worst-unmitigated dependency you identified failed. Use "~X%" if exact numbers are unavailable, "<1%" for de minimis, "unclear" only if A1–A9 left the question genuinely open (in which case grade=unknown is usually correct). Do NOT substitute qualitative phrases like "significant" — give a number or bracket.

GRADE ANCHORS (mapped to DefiScan v1 stages):
- green = Stage 2 equivalent. Failure of any external dependency cannot cause loss of user principal or unclaimed yield. Either there are no material external dependencies, or every critical one has a documented fallback (A6) that keeps users whole. Governance cannot silently introduce new dependencies without an exit window (A9). Impacted TVS under any single-dependency failure: effectively 0.
- orange = Stage 1 equivalent. Failure of some external dependency can cause loss of unclaimed yield, or can materially change expected performance (pause withdrawals, freeze positions, degrade price quality), but cannot cause loss of principal. Committee-based oracles with sanity checks, canonical-only bridges, fallback paths that exist but are incomplete, or governance-mutable dependencies protected by a ≥7-day timelock. Impacted TVS is bounded and recoverable.
- red = Stage 0 equivalent. Failure of an external dependency CAN cause theft or loss of principal. Examples: single-provider oracle with no sanity check or fallback, material TVL on a non-canonical bridge with a guardian multisig, governance can hot-swap oracles or add staking modules with no timelock or exit window, unmitigated keeper-liveness dependency where positions become insolvent if bots stop. Impacted TVS is material.
- unknown = checklist incomplete after inspecting source + verified contracts. Prefer unknown over guessing when A1/A6 could not be reconstructed. RESERVE unknown for cases where the CORE ARCHITECTURE itself is unverifiable — not for cases where you merely cannot enumerate every per-market dependency in a multi-market protocol. If the core router/factory/oracle architecture is verifiable on-chain and you can determine whether the core requires external dependencies, grade the architecture even when an exhaustive per-market external-dependency census is infeasible. Acknowledge the per-market gap in unknowns[] but still issue a grade. Refusing to grade a multi-market protocol because you cannot list every SY/vault/market is over-use of unknown; grade the architecture and say so.

PROMPT-META CHECK (per Hard Rule 17): before finalizing, verify the verdict cites concrete contract addresses, docs, or code — not the rubric itself. If your verdict says "the protocol belongs to a category the rubric marks red", rewrite it with the A1–A9 finding that actually justifies the grade, or drop to grade=unknown.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "autonomy",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Open Access 2/3 submitted Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v3
- protocol.name:              Balancer V3
- protocol.chains:            Ethereum, Monad, Arbitrum, Base, Hyperliquid L1, Optimism, Avalanche, Plasma, xDai
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               [
  {
    "chain": "Ethereum",
    "address": "0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6",
    "role": "Vault Authorizer"
  },
  {
    "chain": "Ethereum",
    "address": "0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f",
    "role": "Balancer DAO Governance (Timelock)"
  },
  {
    "chain": "Ethereum",
    "address": "0x0EF2e7648A9b354822ed0EE9070A2E37332fd242",
    "role": "Emergency SubDAO"
  },
  {
    "chain": "Ethereum",
    "address": "0xA216683066513673FE033917bD5b11f871c65353",
    "role": "Timelock"
  },
  {
    "chain": "Ethereum",
    "address": "0xa731C23D7c95436Baaae9D52782f966E1ed07cc8",
    "role": "Protocol fee controller"
  },
  {
    "chain": "Ethereum",
    "address": "0xba100000625a3754423978a60c9317c58a424e3D",
    "role": "BAL governance token"
  }
]

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
- https://defipunkd.com/address/1/0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6 (Vault Authorizer)
- https://defipunkd.com/address/1/0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f (Balancer DAO Governance (Timelock))
- https://defipunkd.com/address/1/0x0EF2e7648A9b354822ed0EE9070A2E37332fd242 (Emergency SubDAO)
- https://defipunkd.com/address/1/0xA216683066513673FE033917bD5b11f871c65353 (Timelock)
- https://defipunkd.com/address/1/0xa731C23D7c95436Baaae9D52782f966E1ed07cc8 (Protocol fee controller)
- https://defipunkd.com/address/1/0xba100000625a3754423978a60c9317c58a424e3D (BAL governance token)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: OPEN-ACCESS

Evaluate who is allowed to use the protocol and whether any of that permission is granted off-chain.

Scope: this slice is about ADMISSION — who can enter, exit, or transact. Operator LIVENESS (what breaks if keepers/oracles go offline) is assessed in the dependencies slice and is out of scope for the grade here. You may note operator dependencies as context, but do not let "the protocol halts if operator X disappears" drive the access grade on its own; that belongs in dependencies. Source verification / contract verification on block explorers is assessed in the verifiability slice and is out of scope here — do NOT let "contract is unverified" drive the access grade.

Framing: the smart contracts are the access layer; frontends are UX. A permissionless contract is reachable by any client (SDK, third-party UI, aggregator, wallet integration). Frontend ToS, IP geo-blocking, and wallet screening are publisher policies on one specific client — they are reported as context but do NOT determine the grade. The grade hinges on (1) what the contract itself permits, and (2) whether the protocol is practically reachable without the official publisher's cooperation.

Meta-check before finalizing: if your verdict cites phrases from this prompt as evidence ("the protocol meets the 'credible alternatives' condition", "this fits the 'documented fallback' rule"), redo the verdict. The prompt describes the rubric; evidence must come from the protocol. A verdict should cite what the protocol does, not what the rubric says.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. Whitelist / allowlist modifiers in user-facing entry points. Grep for "onlyWhitelisted", "onlyRole", "allowlist", "isAccredited", "isKYCed". Note which functions are gated and who can add/remove from the list.
- A2. Off-chain operators in the admission path: keepers, sequencers, privileged relayers, oracle posters whose approval is required to admit a user action (not just to keep the protocol live). For each, identify whether the role is held by a single operator, a permissioned committee, or is permissionless. Enumerate per user-facing function class (deposit vs withdraw-request vs claim-finalized vs transfer) which ones require operator approval to be admitted, and which ones admit users unconditionally. A function whose placement is unconditional but whose downstream settlement depends on an operator is an admission-permissionless function — flag the liveness dependency as context and defer its grading weight to the dependencies slice.
- A3. Frontend restrictions on the official interface — record as context, not as a grade lever. Distinguish:
    - A3-passive: boilerplate ToS clauses (sanctions attestation, restricted-territory self-certification, VPN-circumvention prohibition, "comply with applicable law" eligibility, age of majority).
    - A3-active: runtime enforcement — IP-based geo-blocking, wallet-address screening against a sanctions oracle (Chainalysis, TRM, Elliptic), KYC wall, rendering-blocking jurisdiction banner.
  Record findings under the correct tier. Quote ToS text or banner text in evidence[].shows. These findings populate the headline and rationale but do NOT move the grade by themselves; the grade is set by A1, A2, and the A3b path check below.
- A3b. Independent access paths (the operative grade input). Enumerate paths that do not require the official publisher's cooperation:
    - Published SDK / library / CLI for direct contract interaction.
    - Third-party frontends operated by separate legal entities.
    - Wallet-integrated access (MetaMask Swaps, Safe apps, etc.).
    - DEX / lending / yield aggregators that route through the contracts.
  Record at least one concrete link per path that exists. The protocol does NOT have to self-document these — the test is existence, not UX cost. An A3b-i redistribution of the official UI bound by the same ToS does NOT count as an independent path.
- A4. Sanctions / compliance tooling at the contract level: does the protocol check addresses against OFAC lists or similar on-chain blocklists in the contract itself? (Frontend-only screening belongs in A3.)
- A5. Differentiate read access vs write access: many protocols are read-permissionless (anyone can view state) but write-gated (only certain addresses can deposit/borrow). Record both.
- A6. ToS / Legal links: locate them on the website and produce a VERBATIM quote of any jurisdictional, sanctions, or eligibility clause in evidence[].shows. If you cannot extract the clause text verbatim (SPA render failure, paywall, dead link, etc.), do NOT paraphrase or infer from general knowledge — record the ToS URL in unknowns[] with the reason extraction failed. Assertions about ToS content without a verbatim quote will be downweighted by reviewers.

Then write the steel-man section per Hard Rule 11.

Grade rules (admission-focused; liveness concerns belong in dependencies; source verification belongs in verifiability):
- green   = no contract-level whitelist/KYC on user entry/exit; no operator approval required to admit a user action; AND at least one independent A3b path exists (published SDK, third-party frontend, wallet integration, or aggregator routing). Frontend ToS posture and A3-active enforcement on the official UI do NOT block green when contracts are permissionless and an independent path exists — they are reported as context.
- orange  = contracts admit users unconditionally, BUT the protocol is operationally captured by the official publisher: no published SDK, no third-party frontend, no wallet integration, no aggregator routing. The contract is theoretically open but practically reachable only through the official UI. Also applies when admission requires approval from a permissioned committee that is governance-managed with a documented replacement procedure.
- red     = contract-level whitelist / KYC on user entry/exit, OR admission of a core user action requires approval from a single privileged operator or a small committee with no documented replacement procedure, OR enforces an on-chain blocklist updatable by a single party.
- unknown = checklist incomplete after checking the sources above.

Default-grade guidance: when contracts are fully permissionless AND any A3b independent path exists, the default grade is green regardless of frontend ToS or A3-active enforcement on the official UI. Frontend geo-blocking, sanctions-oracle wallet screening, and ToS sanctions clauses are publisher policies on one client and are reported in findings/headline as context, not as grade levers. To grade orange on operational-capture grounds, the auditor must affirmatively show that ALL independent paths are absent or also gated.

Guideline on committees: where admission depends on a multi-operator committee, the relevant axes are (a) set size, (b) whether replacement/rotation is governed on-chain, (c) whether the replacement procedure is publicly documented. A large set with on-chain governance replacement should not be graded as a single-party operator even if rotation is not instantaneous. A small set with informal replacement should be treated as a single-party operator.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "open-access",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Audit all 5 dimensions · one prompt Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v3
- protocol.name:              Balancer V3
- protocol.chains:            Ethereum, Monad, Arbitrum, Base, Hyperliquid L1, Optimism, Avalanche, Plasma, xDai
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               [
  {
    "chain": "Ethereum",
    "address": "0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6",
    "role": "Vault Authorizer"
  },
  {
    "chain": "Ethereum",
    "address": "0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f",
    "role": "Balancer DAO Governance (Timelock)"
  },
  {
    "chain": "Ethereum",
    "address": "0x0EF2e7648A9b354822ed0EE9070A2E37332fd242",
    "role": "Emergency SubDAO"
  },
  {
    "chain": "Ethereum",
    "address": "0xA216683066513673FE033917bD5b11f871c65353",
    "role": "Timelock"
  },
  {
    "chain": "Ethereum",
    "address": "0xa731C23D7c95436Baaae9D52782f966E1ed07cc8",
    "role": "Protocol fee controller"
  },
  {
    "chain": "Ethereum",
    "address": "0xba100000625a3754423978a60c9317c58a424e3D",
    "role": "BAL governance token"
  }
]

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
- https://defipunkd.com/address/1/0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6 (Vault Authorizer)
- https://defipunkd.com/address/1/0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f (Balancer DAO Governance (Timelock))
- https://defipunkd.com/address/1/0x0EF2e7648A9b354822ed0EE9070A2E37332fd242 (Emergency SubDAO)
- https://defipunkd.com/address/1/0xA216683066513673FE033917bD5b11f871c65353 (Timelock)
- https://defipunkd.com/address/1/0xa731C23D7c95436Baaae9D52782f966E1ed07cc8 (Protocol fee controller)
- https://defipunkd.com/address/1/0xba100000625a3754423978a60c9317c58a424e3D (BAL governance token)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: CONTROL

Evaluate who can change the protocol's rules, how fast, and how broadly.

(Step 0 capability probe and the off-chain-only fallback live in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[])

- **C1.** For each address you assess: who is the contract owner / admin / pendingAdmin / governor — read these via the block explorer's "Read Contract" tab OR `https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=owner` (BARE method names: `&method=owner`, `&method=admin`, `&method=pendingOwner`, `&method=governor`). For Safes use `/api/safe/owners?chainId=<id>&address=0x...`. When a protocol has multiple major versions deployed (v2/v3/v4), perform C1 reads on the NEWEST deployment separately — newer deployments often have weaker control surfaces than the legacy core.
- **C2.** Upgrade mechanism: transparent proxy / UUPS / Beacon / Diamond / immutable. Identify the proxy admin address. Check upgradeability of GOVERNANCE contracts too — a Governor / Aragon Voting / OZ Governor is often itself a proxy whose admin is the Timelock. Asymmetry: when fund-holding cores are immutable AND governance has no admin path that reaches them, an upgradable Governor/Timelock is T3-only and must NOT drag the verdict below green on that basis alone (see grade rules and the "immutable cores" caveat). Only call upgradability "mixed" if you can name a concrete function on the upgradable surface that reaches T1 or T2 on user funds.
- **C3.** EXECUTION PATH (enumerate every stage, in order, with delays in seconds). The operative path is usually a chain (voting → scheduler → timelock → executor; or governor → queue → execute; or Aragon Voting → DualGovernance → EmergencyProtectedTimelock → AdminExecutor). For each stage, record (a) the contract address, (b) the delay constant name + value in seconds, (c) the URL you read it from (block-explorer Read Contract OR `/api/contract/read?...&method=MIN_DELAY&block=<n>`). Do NOT stop at the first timelock-shaped contract — if its admin is itself called by another contract, keep walking. The grading delay is the SUM OF DELAYS ON THE UNCONTESTED FAST PATH (shortest time a proposal with no opposition can go from submission to executable). Dynamic / contested extensions (veto signaling, rage quit, escrow delay) are modifiers, not the basis — note them separately.
- **C4.** Enumerate EVERY multisig with reachable control — main proxy admin, emergency activation, emergency execution, reseal / pause, gate-seal committees, tiebreaker, per-module admins. For each Safe, fetch threshold + owners + version via `/api/safe/owners?chainId=<id>&address=0x...` (response includes raw eth_call data, so the URL is citable evidence). Enumerate ops/council/incentives multisigs even when off the upgrade path — record their scope so a reader can see they are NOT on the upgrade path. For each: (a) address, (b) threshold / total signers, (c) signer identities classified as insider (team, paid auditors under ongoing engagement, mandated service providers) vs non-insider (independent community members, unaffiliated researchers), (d) the specific power held (upgrade, pause, parameter, etc.).
- **C5.** On-chain governance: Governor / GovernorBravo / OZ Governor / Aragon Voting with token-weighted voting? Record proposal threshold, voting period, quorum, and the timelock delay between queue and execute. Every numeric constant must come from a Read Contract call you can link to, or be in unknowns[] with the C-code. If votingDelay / votingPeriod are denominated in BLOCKS, convert to seconds at the chain's CURRENT block time (Ethereum mainnet ≈ 12s post-Merge, not the 15s in older Compound/Bravo deployments) — cite both block count and converted seconds.
- **C6.** EMERGENCY POWERS: separate emergency-pause / guardian role with a different time cap or different actor than the main upgrade authority? Record it explicitly.
- **C7.** POWER TIER (blast radius). For each privileged path in C3–C6, classify the WORST thing that path can do, choosing the highest applicable tier. Cite the specific function name and any on-chain bound — tier claims without a named function are unsupported.
    - **T1 — FUND-CRITICAL**: replace implementation of contracts holding user funds; change AMM math / accounting / collateral logic; mint unbacked debt or shares; pause withdrawals; drain user-fund treasury; change oracle to attacker-controlled source; replace upgrade admin with EOA.
    - **T2 — ECONOMICALLY MATERIAL**: change fee parameters within bounded ranges; redirect protocol fees; add/remove markets / collateral types; bounded inflation or token mint within hard-capped schedule; spend protocol-owned (non-user) treasury.
    - **T3 — GOVERNANCE-INTERNAL**: change voting rules, quorum, voting period, proposal threshold; upgrade the Governor itself; rotate Timelock admin; mint governance tokens within a capped annual schedule.
    - **T4 — OPERATIONAL**: incentives distribution, grants, ENS / frontend canonicalization, deployment coordination, periphery router deprecation.
  The grade is set by the HIGHEST tier reachable on the uncontested fast path, not the median. State the tier and the binding function in the verdict.

### Read Contract discipline (applies to C3, C4, C5)

Every numeric constant cited (timelock delays, voting periods, multisig thresholds, quorum percentages) must come from EITHER (a) a block-explorer Read Contract URL, OR (b) a DeFiPunkd `/api/contract/read` or `/api/safe/owners` URL (preferred with `&block=<n>` for content-addressed evidence), OR (c) an unknowns[] entry with the C-code. Docs / blog posts are corroboration only — they cannot be the sole citation for a value that is also readable on-chain.

### Off-chain-only substitute hierarchy (when grading_basis="off-chain-only" — see preamble Rule 16)

When on-chain reads were genuinely unreachable this run, eligible off-chain substitutes in priority order:
1. Linked audit PDFs (admin roles, multisig members, timelock delays usually enumerated).
2. Governance forum posts that quote constants from a successful on-chain proposal (cite post URL + linked execution-tx URL).
3. Official protocol docs pages with named addresses and roles (must be on a domain owned by the protocol).
4. GitHub README / SECURITY.md / governance/*.md at a pinned commit SHA.

Forbidden substitutes: third-party blog posts, X / Twitter threads, search-result snippets, model memory. Required degradation: any C-code citing a numeric constant from docs/forum/audit prose ONLY must also carry an `unknowns[]` entry with `-offchain` suffix noting "value not re-read on-chain in this run; corroboration only".

### Grade rules (apply the timelock bar conditional on the highest C7 tier reachable on the fast path)

Security Council standard (used below): a multisig qualifies as "Security Council" only if ALL of: ≥7 signers, ≥51% threshold, ≥50% non-insider signers, every signer publicly announced. Failing any criterion = NOT a Security Council, regardless of signer reputation.

- **green**: highest reachable tier is T3 or T4 regardless of timelock; OR T2 reachable with uncontested-fast-path delay ≥7 days; OR T1 reachable only via immutable contracts (T1 is unreachable); OR T1 reachable with uncontested-fast-path delay ≥7 days combined with a Security Council multisig; OR T1 reachable with uncontested-fast-path delay ≥7 days through active on-chain governance with broad token distribution.
- **orange**: T2 reachable with uncontested-fast-path delay >0 but <7 days; OR T1 reachable with uncontested-fast-path delay >0 but <7 days; OR a multisig failing one or more Security Council criteria sits on a T1/T2 path; OR unclear upgrade authority on a T1/T2 path; OR governance with very short timelock or low quorum on a T1/T2 path.
- **red**: T1 reachable with no timelock by a single EOA or 2-of-3 multisig; OR a T1 upgrade admin that is not a smart contract you can audit.
- **unknown**: completed the checklist but still cannot determine the upgrade authority OR cannot classify the highest tier reachable on the main contracts.

Tiering caveats:
- "Bounded" must be enforced ON-CHAIN to count as T2. A function that sets fees with no upper-bound check is T1 — cite the bound check.
- Recurring T2 economic extraction (e.g. fee redirect with no rate limit) approaches T1 over time. A single proposal that can permanently redirect all future revenue is T1.
- T3 assumes the governance contract cannot itself authorize a T1/T2 action without going through the same timelock. If governance can self-upgrade to bypass the timelock, T3 collapses into T1.
- Do not downgrade tier by hand-waving ("realistically governance would never…"). Tier on what the contract permits, not what feels likely.

Notes:
- **Dynamic / dual-governance timelocks** (Lido, Compound escrow veto): the rubric grades on the uncontested path because that is the path most upgrades take. A dynamic extension that fires only under stake-weighted opposition is a real protection — name it in the green steel-man, but it does not lift an orange fast path into green; state the tension in the verdict.
- **Immutable cores with upgradable governance** (Uniswap-style): if fund-holding contracts are immutable and have no admin-reachable function moving / freezing / re-routing user funds, the highest reachable tier on the upgrade path is T3 — green regardless of timelock. Don't grade this orange just because the Governor is a proxy — that's a C2 fact, not a downgrade. Downgrade only applies if you can cite a concrete function on the upgradable surface that reaches T1 or T2 (privileged hook, upgradable factory controlling fund-routing, fee-switch redirecting protocol revenue without bound).
- **The 7-day bar** reflects the exit-window standard — users need notice after a queued upgrade to withdraw if they disagree. The ability-to-exit slice grades the exit side; this slice grades the delay side; both must hold for users to actually benefit from the delay.

---

### Slice: ABILITY-TO-EXIT

Evaluate whether users can withdraw their funds on their own terms, even under adversarial admin conditions.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- E1. Enumerate every user-facing exit function in the main contracts: withdraw, redeem, burn, requestWithdrawal, claim, exit, etc. List them by name. Do NOT treat the contract as a monolith.
- E2. For EACH exit function in E1: identify its access modifiers and any pause guards (e.g. _checkResumed, whenNotPaused, onlyRole). Functions that gate REQUEST PLACEMENT often differ from functions that CLAIM FINALIZED FUNDS — check both separately.
- E3. For each pause guard: identify the role holder (which address holds PAUSE_ROLE / GUARDIAN / etc.) and the maximum pause duration. Specifically check whether PAUSE_INFINITELY (or equivalent uncapped pause) is callable, and which actor can call it (single multisig vs governance vote). For role-holder reads use https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=hasRole&args=0x...,0x... or &method=getRoleAdmin&args=0x.... For "is currently paused" checks use &method=paused or &method=isPaused&args=<resume-code>. Use the BARE method name (no parens). Cite the URL with &block=<n> in evidence[].
- E4. EMERGENCY vs GOVERNANCE pause distinction: many protocols have a fast-acting emergency pause capped at N days and a slow governance pause that can be indefinite. Record both paths separately if present, with their time caps and actor classes.
- E5. Queued redemption: documented maximum queue duration, daily withdrawal caps, whether the queue itself is pausable.
- E6. Forced-exit / escape-hatch / permissionless emergency-exit mechanism for adversarial-admin scenarios.
- E7. Frontend dependency: confirm exit functions are directly callable on-chain (e.g. via Etherscan write tab or a generic wallet) without the project's frontend.

Then write the steel-man section per Hard Rule 11. Common red-vs-orange tension on this slice: indefinite pause exists (suggests red) BUT the realistic emergency path is time-capped AND claims of already-finalized exits are not pause-gated (suggests orange). Resolve this by stating who can do what for how long, not by stopping at the worst-case sentence.

Grade rules:
- green   = permissionless exit; pause is either absent, narrowly scoped to clearly-described emergencies with auto-expiry, or capped at ≤7 days; no frontend dependency for exit; claims of already-finalized exits are not pause-gated under any path.
- orange  = pausable with broad scope OR indefinite pause is reachable only through governance vote (not unilateral admin action), OR queued redemption with documented max > 7 days, OR claims-of-finalized are exempt but new-request placement can be paused indefinitely by governance.
- red     = exit requires admin signature, OR ANY actor (including governance) can pause CLAIMS of finalized exits indefinitely, OR there is no on-chain exit function at all (purely custodial), OR pause is held by a single EOA / 2-of-3 multisig with no time cap.
- unknown = checklist incomplete after checking the sources above.

---

### Slice: AUTONOMY

Evaluate this protocol's autonomy: can a failure of anything outside its own contracts cause theft or loss of user principal, loss of unclaimed yield, or materially change the protocol's expected performance? "Autonomy" is not "has zero external touchpoints" — it is "external failures are either survivable or recoverable without user loss". This slice adapts DefiScan v1's Autonomy dimension; dependencies are one of several criteria, not the whole frame.

GUARDRAILS (read before grading):
- Category alone (Liquid Staking, Bridge, RWA Lending, Restaking, …) does NOT force a grade. A category is a hint about where to look; the grade must come from the concrete A1–A9 findings below.
- Base-chain consensus (Ethereum PoS, the chain's validator set, the canonical Deposit Contract at 0x00000000219ab540356cBB839Cbe05303d7705Fa) is the SUBSTRATE, not a dependency, for any protocol deployed on that chain. Do not list "depends on Ethereum" as a finding.
- Oracles or other integrations used by DOWNSTREAM protocols that happen to read this protocol's token (e.g. Chainlink stETH/USD consumed by Aave) are NOT this protocol's dependencies. Count only what THIS protocol's contracts call or trust on-chain. If a feed or contract appears in the protocol's docs only as reference material for third-party integrators, EXCLUDE IT ENTIRELY — do not log it as a finding even with a "peripheral" or "referenced only" caveat.
- "Upgradeable admin can change things" belongs to the CONTROL slice; only count it here when the upgrade surface lets governance silently swap an external dependency (see A9).
- Underlying-asset risk in opt-in, isolated markets is NOT autonomy-red on its own. When a protocol wraps third-party yield-bearing assets (LSTs, LRTs, ERC-4626 vaults, lending receipts, restaked tokens) into per-market silos that users explicitly choose, a failure of one underlying does NOT propagate to other markets and is risk the user opted into per-market. Record it under A4 with depth and propagation scope, but do not let it alone drive a red verdict — red requires an external dependency that cross-cuts the protocol or that the user did not opt into at deposit time. A failure mode that is "if the LRT you deposited is hacked, your principal in that LRT-backed market is impaired" is the underlying's autonomy story, not this protocol's; grade it on whether THIS protocol introduces additional dependencies on top.
- Sub-module enumeration is mandatory before grading. If the protocol ships distinct product lines or modules (e.g., a v2 core AMM plus a newer perps/funding-rate module, a lending pool plus a separate vault layer, an L1 core plus a cross-chain extension), enumerate each in findings and grade against the WORST module weighted by its share of TVS or its blast radius. A green core does not rescue an orange/red sub-module; conversely, a small red sub-module with capped TVS may bound the overall grade to orange. Name each module by its on-chain factory or router address. If you do not know whether a module exists, that is an unknowns[] entry, not silence. EXPLICIT TVS WEIGHTING: in the verdict, state each module's approximate share of total protocol TVS (use "~X%" if exact figures unavailable; check DeFiLlama or block-explorer balances on the module's main contract) and how that share informs the weighted grade. Format: "Module A holds ~X% of TVS (grade: <g>); Module B holds ~Y% (grade: <g>); weighted overall = <grade> because <reason>." If a red sub-module holds <5% of total TVS and is capped, the overall grade may be orange; if it holds >25%, the overall grade is red. Do not let qualitative reasoning substitute for the percentages — write the numbers.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. External contract calls. Enumerate every external contract the core contracts call or read from (oracles, price feeds, AMM pools, lending pools, staking/deposit contracts, yield wrappers). For each, identify the address, the provider, and what user-facing function of this protocol would break or mis-price if that external contract paused, mis-reported, or behaved adversarially. Grep for "oracle", "aggregator", "getPrice", "latestAnswer", "chainlink", "pyth", "redstone" as a starting point. To verify an oracle is live and what it currently reports, hit https://defipunkd.com/api/contract/read?chainId=<id>&address=<oracle>&method=latestAnswer (or &method=latestRoundData, &method=getPrice, &method=description; use the BARE method name without parens) and cite the URL — the response includes blockNumber/blockHash and rawReturnData, which is stronger evidence than a docs page about which feed "should" be used.
- A2. Off-chain actor committees reporting INTO the protocol. Oracle committees, guardian multisigs, DAO-selected validator sets acting as protocol reporters, exit-bus signers, fraud-proof challengers. For each, record committee size, quorum, who picks the members, and what mis-reporting could do (mint, burn, finalize withdrawal, freeze). Distinguish this from governance admins (control slice). NOTE ON STAKING PROTOCOLS: validator slashing and node-operator misbehavior are properties of the base-chain substrate, NOT external dependencies, when the operator set is diversified enough that a coordinated failure caps at <5% principal loss. Count validator/operator risk under A2 ONLY if the operator set is small, non-diversified, or lacks bonding / slashing-insurance / diversification mitigations. A curated set of 30+ independent operators with documented diversification falls under the mitigated path; a 3-operator LST does not. Do not cite the protocol's own risk disclosure as evidence that operator failure = principal loss unless you also check the diversification and bond mitigations.
- A3. Bridge / cross-chain messaging dependencies. Only count bridges that carry material TVL or are required for a core user flow. For each, name the bridge operator (canonical L1↔L2, LayerZero, Wormhole, Axelar, custom multisig), the trust model (canonical, optimistic, light-client, guardian set), and what fraction of TVL or users ride on it. "wstETH exists on 15 chains" is not a finding unless material TVL sits there. Before listing any non-primary chain deployment as a dependency, verify it is still operational as of analysis_date — retired or sunset deployments (e.g., Lido-on-Terra, Lido-on-Solana) belong in unknowns[] or should be omitted, never cited as a current dependency.
- A4. Nested collateral / restaking chains. For restaking / LRT / receipt-of-receipt designs: record the depth of the collateral chain, every actor with slashing or freezing power at each level, and whether a failure N levels deep propagates to user principal here.
- A5. Fork lineage (silent check). If DeFiLlama's forkedFrom is non-empty, record it as one finding and move on. If empty, do NOT add a placeholder finding; it adds noise.
- A6. Fallback mechanisms and circuit breakers. What catches an external failure? Sanity-check contracts on oracle reports, rebase bounds, pause paths triggered by bad prices, second-opinion oracles, max-per-block throughput caps, withdrawal queues that absorb bad reports. Record which A1–A4 risks are mitigated by which fallback, and which are unmitigated. For EACH fallback you cite, state its activation status explicitly: (i) LIVE and enforcing on-chain today, (ii) DEPLOYED but not yet wired / activated (e.g., interface exists but the address is zero or the role is unassigned), or (iii) DOCUMENTED / PROPOSED only (forum post, LIP draft, audit pending). Only (i) counts as mitigation for the grade. A fallback in state (ii) or (iii) should be noted but must not reduce the risk in your steel-man or verdict. If you cannot determine activation status, add an unknowns[] entry rather than assume it is live.
- A7. Sequencer / L1-liveness dependency BEYOND the base-chain substrate. SCOPE — sequencer risk only counts here when the protocol IS its own L2/L3 appchain or app-rollup, where the sequencer is part of the protocol's own stack and a freeze is a protocol-level outage. A protocol permissionlessly deployed on a third-party L2 inherits that L2's sequencer as substrate, not an A7 dependency. Record the sequencer/DA trust model when A7 applies.
- A8. Keeper / relayer / off-chain bot liveness. Protocols that need permissionless-but-necessary off-chain actors (liquidation bots, auto-compounders, deposit relayers, intent solvers). Record whether the role is permissionless, what degrades if nobody runs it (yield paused, bad debt accumulates, positions go stale), and whether the failure mode is graceful or catastrophic.
- A9. Governance-mutable dependency surface. Can an admin or DAO action silently INTRODUCE a NEW EXTERNAL dependency — swap the oracle address to a different provider, register a new staking module that calls an untrusted contract, add a new bridge, route SY through a new external vault — without an exit window for users? Check the upgrade / router / module-registry contracts. Answer: which EXTERNAL dependencies are governance-mutable, who holds that power, and whether there is a timelock or exit window. SCOPE LIMIT — read carefully: A9 is about the *external dependency surface*, not the upgrade surface in general. "The proxyAdmin / EOA can upgrade the router implementation to arbitrary bytecode" is a CONTROL-slice finding (admin can rug), NOT an autonomy-A9 finding. A9 fires only when the upgrade specifically swaps out or adds a contract that THIS protocol calls or trusts (e.g., changing the oracle address from Chainlink to a malicious feed, registering a new SY adapter that points to a third-party vault, redirecting a bridge endpoint). If the only finding is "admin can change implementation," do not log it under A9 and do not let it drive the autonomy grade — note it under control instead. The autonomy-relevant version of the same upgrade key is "admin can swap [specific external address X] without timelock"; that requires identifying the specific external dependency that becomes mutable.

STEEL-MAN (per Hard Rule 13): write one-sentence strongest arguments for red, orange, and green using the A1–A9 findings.

IMPACTED TVS ESTIMATE: the headline MUST include a rough impacted-TVS figure — the fraction of protocol TVS that could be lost or frozen if the worst-unmitigated dependency you identified failed. Use "~X%" if exact numbers are unavailable, "<1%" for de minimis, "unclear" only if A1–A9 left the question genuinely open (in which case grade=unknown is usually correct). Do NOT substitute qualitative phrases like "significant" — give a number or bracket.

GRADE ANCHORS (mapped to DefiScan v1 stages):
- green = Stage 2 equivalent. Failure of any external dependency cannot cause loss of user principal or unclaimed yield. Either there are no material external dependencies, or every critical one has a documented fallback (A6) that keeps users whole. Governance cannot silently introduce new dependencies without an exit window (A9). Impacted TVS under any single-dependency failure: effectively 0.
- orange = Stage 1 equivalent. Failure of some external dependency can cause loss of unclaimed yield, or can materially change expected performance (pause withdrawals, freeze positions, degrade price quality), but cannot cause loss of principal. Committee-based oracles with sanity checks, canonical-only bridges, fallback paths that exist but are incomplete, or governance-mutable dependencies protected by a ≥7-day timelock. Impacted TVS is bounded and recoverable.
- red = Stage 0 equivalent. Failure of an external dependency CAN cause theft or loss of principal. Examples: single-provider oracle with no sanity check or fallback, material TVL on a non-canonical bridge with a guardian multisig, governance can hot-swap oracles or add staking modules with no timelock or exit window, unmitigated keeper-liveness dependency where positions become insolvent if bots stop. Impacted TVS is material.
- unknown = checklist incomplete after inspecting source + verified contracts. Prefer unknown over guessing when A1/A6 could not be reconstructed. RESERVE unknown for cases where the CORE ARCHITECTURE itself is unverifiable — not for cases where you merely cannot enumerate every per-market dependency in a multi-market protocol. If the core router/factory/oracle architecture is verifiable on-chain and you can determine whether the core requires external dependencies, grade the architecture even when an exhaustive per-market external-dependency census is infeasible. Acknowledge the per-market gap in unknowns[] but still issue a grade. Refusing to grade a multi-market protocol because you cannot list every SY/vault/market is over-use of unknown; grade the architecture and say so.

PROMPT-META CHECK (per Hard Rule 17): before finalizing, verify the verdict cites concrete contract addresses, docs, or code — not the rubric itself. If your verdict says "the protocol belongs to a category the rubric marks red", rewrite it with the A1–A9 finding that actually justifies the grade, or drop to grade=unknown.

---

### Slice: OPEN-ACCESS

Evaluate who is allowed to use the protocol and whether any of that permission is granted off-chain.

Scope: this slice is about ADMISSION — who can enter, exit, or transact. Operator LIVENESS (what breaks if keepers/oracles go offline) is assessed in the dependencies slice and is out of scope for the grade here. You may note operator dependencies as context, but do not let "the protocol halts if operator X disappears" drive the access grade on its own; that belongs in dependencies. Source verification / contract verification on block explorers is assessed in the verifiability slice and is out of scope here — do NOT let "contract is unverified" drive the access grade.

Framing: the smart contracts are the access layer; frontends are UX. A permissionless contract is reachable by any client (SDK, third-party UI, aggregator, wallet integration). Frontend ToS, IP geo-blocking, and wallet screening are publisher policies on one specific client — they are reported as context but do NOT determine the grade. The grade hinges on (1) what the contract itself permits, and (2) whether the protocol is practically reachable without the official publisher's cooperation.

Meta-check before finalizing: if your verdict cites phrases from this prompt as evidence ("the protocol meets the 'credible alternatives' condition", "this fits the 'documented fallback' rule"), redo the verdict. The prompt describes the rubric; evidence must come from the protocol. A verdict should cite what the protocol does, not what the rubric says.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. Whitelist / allowlist modifiers in user-facing entry points. Grep for "onlyWhitelisted", "onlyRole", "allowlist", "isAccredited", "isKYCed". Note which functions are gated and who can add/remove from the list.
- A2. Off-chain operators in the admission path: keepers, sequencers, privileged relayers, oracle posters whose approval is required to admit a user action (not just to keep the protocol live). For each, identify whether the role is held by a single operator, a permissioned committee, or is permissionless. Enumerate per user-facing function class (deposit vs withdraw-request vs claim-finalized vs transfer) which ones require operator approval to be admitted, and which ones admit users unconditionally. A function whose placement is unconditional but whose downstream settlement depends on an operator is an admission-permissionless function — flag the liveness dependency as context and defer its grading weight to the dependencies slice.
- A3. Frontend restrictions on the official interface — record as context, not as a grade lever. Distinguish:
    - A3-passive: boilerplate ToS clauses (sanctions attestation, restricted-territory self-certification, VPN-circumvention prohibition, "comply with applicable law" eligibility, age of majority).
    - A3-active: runtime enforcement — IP-based geo-blocking, wallet-address screening against a sanctions oracle (Chainalysis, TRM, Elliptic), KYC wall, rendering-blocking jurisdiction banner.
  Record findings under the correct tier. Quote ToS text or banner text in evidence[].shows. These findings populate the headline and rationale but do NOT move the grade by themselves; the grade is set by A1, A2, and the A3b path check below.
- A3b. Independent access paths (the operative grade input). Enumerate paths that do not require the official publisher's cooperation:
    - Published SDK / library / CLI for direct contract interaction.
    - Third-party frontends operated by separate legal entities.
    - Wallet-integrated access (MetaMask Swaps, Safe apps, etc.).
    - DEX / lending / yield aggregators that route through the contracts.
  Record at least one concrete link per path that exists. The protocol does NOT have to self-document these — the test is existence, not UX cost. An A3b-i redistribution of the official UI bound by the same ToS does NOT count as an independent path.
- A4. Sanctions / compliance tooling at the contract level: does the protocol check addresses against OFAC lists or similar on-chain blocklists in the contract itself? (Frontend-only screening belongs in A3.)
- A5. Differentiate read access vs write access: many protocols are read-permissionless (anyone can view state) but write-gated (only certain addresses can deposit/borrow). Record both.
- A6. ToS / Legal links: locate them on the website and produce a VERBATIM quote of any jurisdictional, sanctions, or eligibility clause in evidence[].shows. If you cannot extract the clause text verbatim (SPA render failure, paywall, dead link, etc.), do NOT paraphrase or infer from general knowledge — record the ToS URL in unknowns[] with the reason extraction failed. Assertions about ToS content without a verbatim quote will be downweighted by reviewers.

Then write the steel-man section per Hard Rule 11.

Grade rules (admission-focused; liveness concerns belong in dependencies; source verification belongs in verifiability):
- green   = no contract-level whitelist/KYC on user entry/exit; no operator approval required to admit a user action; AND at least one independent A3b path exists (published SDK, third-party frontend, wallet integration, or aggregator routing). Frontend ToS posture and A3-active enforcement on the official UI do NOT block green when contracts are permissionless and an independent path exists — they are reported as context.
- orange  = contracts admit users unconditionally, BUT the protocol is operationally captured by the official publisher: no published SDK, no third-party frontend, no wallet integration, no aggregator routing. The contract is theoretically open but practically reachable only through the official UI. Also applies when admission requires approval from a permissioned committee that is governance-managed with a documented replacement procedure.
- red     = contract-level whitelist / KYC on user entry/exit, OR admission of a core user action requires approval from a single privileged operator or a small committee with no documented replacement procedure, OR enforces an on-chain blocklist updatable by a single party.
- unknown = checklist incomplete after checking the sources above.

Default-grade guidance: when contracts are fully permissionless AND any A3b independent path exists, the default grade is green regardless of frontend ToS or A3-active enforcement on the official UI. Frontend geo-blocking, sanctions-oracle wallet screening, and ToS sanctions clauses are publisher policies on one client and are reported in findings/headline as context, not as grade levers. To grade orange on operational-capture grounds, the auditor must affirmatively show that ALL independent paths are absent or also gated.

Guideline on committees: where admission depends on a multi-operator committee, the relevant axes are (a) set size, (b) whether replacement/rotation is governed on-chain, (c) whether the replacement procedure is publicly documented. A large set with on-chain governance replacement should not be graded as a single-party operator even if rotation is not instantaneous. A small set with informal replacement should be treated as a single-party operator.

---

### Slice: VERIFIABILITY

Evaluate whether an outsider can independently confirm what the deployed code does.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- V1. For each address you assess: is the bytecode verified on the chain's block explorer? Record the "Contract Source Code Verified" indicator. https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x... returns this as a top-level "verified" boolean plus "abiSource" ("etherscan" / "sourcify") and an inline ABI — useful when the explorer page is rate-limited. If the contract is a proxy, verify BOTH the proxy contract AND the current implementation contract. The same /api/contract/abi response auto-resolves proxies and includes a "proxy.implementation" address when present, so one call covers V1 + V6 in one shot. An explorer "Similar Match" on a well-known proxy pattern (Aragon AppProxyUpgradeable, ERC1967Proxy, OssifiableProxy, OZ TransparentUpgradeableProxy) is expected for that pattern and does NOT count as a verification gap on its own — what matters is that the implementation is independently verified.
- V2. Source-to-repo correspondence: for each verified contract, attempt to find a matching commit in the linked GitHub repos. Record evidence[].commit on a match. Independent compile/bytecode-match is NOT required for green — a recognized public repo whose structure and file contents correspond to the explorer-visible source is sufficient. If you did not pin a commit SHA or run a bytecode diff, record that plainly in unknowns[] and proceed; it is a scope limit, not a downgrade signal.
- V3. Audit coverage: for each URL in protocol.audit_links, open it and record: auditor name, audit date, the specific contracts / commit in scope. Flag audits that predate the current deployment by >6 months without a follow-up review.
- V4. Auditor recognition: the following firms are broadly recognized in Solidity: Trail of Bits, Zellic, Spearbit, OpenZeppelin, ConsenSys Diligence, Certora (formal verification), Quantstamp, Halborn, Peckshield, Sigma Prime, ChainSecurity, Ackee Blockchain, MixBytes, Statemind. Unknown firms are orange-at-best for any green-grade claim. Name the firm explicitly in evidence[].
- V5. Post-audit drift: compare the most recent in-scope audit(s) against the currently-deployed source, weighted by what each contract does and by what the changes actually contain. SCOPE — drift only downgrades the grade when ALL of the following hold: (i) the drifting contract is fund-custody / settlement / accounting-critical (NOT a peripheral router, lens, quoter, or pure-view contract that holds no balances); (ii) the changes are material — new functions, modified access control, modified accounting, modified fund flow — and not refactors / struct relocations / import reorgs / build or CI fixes / formatting; (iii) no later audit, fix-audit, or differential audit from a recognized firm covers the changed files (audits often pin a pre-fix commit while a follow-up reviews the delta — match by file scope, not by commit-hash equality). When you cite drift as a downgrade reason, name the specific behavior change (function added, role granted, accounting formula altered) — "N commits ahead" or "+X/-Y LOC" alone is not evidence of material drift. If you have not sampled the diff content (e.g. via the GitHub compare view or the top commits in the window), record drift as an unknowns[] entry rather than auto-downgrading; commit-count and LOC are starting signals, not findings.
- V6. Implementation vs proxy: a verified proxy with an unverified implementation is effectively unverified. State whether the implementation is verified separately.

EVIDENCE DISCIPLINE (read before writing findings[]):
- Do not assert a specific deploy-commit SHA, bytecode equivalence, or "identical to audited commit" unless you actually fetched the artifact that shows it (e.g., a deployed-addresses JSON you opened, an explorer page you read). Inferred or plausible matches belong in unknowns[], never in findings[] or evidence[].
- Evidence[] entries must correspond to pages/files you actually retrieved this run. A URL you did not open is not evidence.

Then write the steel-man section per Hard Rule 11.

Grade rules:
- green   = deployed bytecode verified on the explorer (proxy AND implementation if proxied; "Similar Match" on a standard proxy pattern is fine per V1), a public source repo exists whose contents correspond to the explorer-visible source, AND ≥1 audit from a recognized firm covering the currently-deployed contracts (≤6 months of drift OR drift was re-audited). A missing local compile-match is not a downgrade — record it in unknowns[] and still grade green if the other conditions hold.
- orange  = verified but with visible drift from the public repo, OR audit scope is stale relative to deployment, OR only minor / unknown-firm audits exist, OR only some of the main contracts are verified, OR proxy verified but implementation only partially verified.
- red     = unverified bytecode (or verified proxy with unverified implementation), OR no audit in protocol.audit_links, OR no public repo.
- unknown = reserved for when the protocol's verifiability posture genuinely cannot be assessed (e.g., explorer and repo both inaccessible for this protocol). Do NOT use unknown merely because you, the analyst, could not run a particular check such as a bytecode diff — that goes in unknowns[] while the grade is still assigned from the evidence you do have.

---

### JSON output contract

Return exactly one JSON array inside a single ```json fenced block. The array MUST contain exactly five objects, one per risk slice, in this exact order: "control", "ability-to-exit", "autonomy", "open-access", "verifiability". Do not include the discovery slice.

Each object has the same shape as a normal slice submission:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "<one of: control | ability-to-exit | autonomy | open-access | verifiability>",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- Produce one complete object for each of these slices only: control, ability-to-exit, autonomy, open-access, verifiability.
- Reuse the same model, chat_url, snapshot_generated_at, prompt_version, analysis_date, and slug values across all five objects.
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches that object's slice checklist prefix verbatim (C1, E2, AU3, A3b, V4a, …); unknowns[] entries are checklist-coded ("C3: …").
- Wrap the array in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Verifiability tentative 3/3 models agree AI-only 0/3 with chat share link

Audits from Trail of Bits, Spearbit, and Certora exist; Authorizer verified on-chain; but 4 of 5 pinned addresses unconfirmed this run, audit README explicitly flags post-audit contract changes, and no commit SHA pinned

Verdict

Choosing orange because: recognized-firm audit coverage (Spearbit, Certora, Trail of Bits) and the Authorizer's verification are positive signals, but the audits README explicitly acknowledges post-audit contract modifications to the Vault (swap fee accounting in PR #1113), only 1 of 5 pinned addresses had on-chain verification confirmed in this run, no commit SHA was pinned, and the specific scope of the Nov 2024 follow-up review relative to fund-custody changes cannot be confirmed without fetching the audit PDFs — collectively these represent stale-relative-to-deployment audit scope uncertainty that holds the grade below green.

Steelman argument

Steelman argument While three recognized firms audited V3 pre-launch and a Cantina follow-up review was done in Nov 2024, only 1 of 5 pinned contract addresses had verification status confirmed on-chain this run; the audit README flags post-audit changes; no commit SHA was pinned; and the Vault PDF scopes could not be directly verified.

Evidence (6)

V1: The Authorizer contract (0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6) is verified on Etherscan (abiSource: 'etherscan', verified: true) and is not a proxy (proxy: null), so no implementation gap. The remaining 4 pinned addresses (Balancer DAO Governance Timelock, Emergency SubDAO, Timelock, Protocol fee controller) could not have their verification status confirmed in this run due to API fetch failures.
V2: The canonical source repo https://github.com/balancer/balancer-v3-monorepo was identified via search and the repo README confirms it contains Balancer V3 core smart contracts including the Vault. However, GitHub was not directly fetchable this run, so no specific commit SHA could be pinned for source-to-deployed correspondence. Source-to-repo match is plausible but unconfirmed at commit level.
V3: The GitHub audits directory (github.com/balancer/balancer-v3-monorepo/tree/main/audits) was found via search and contains at minimum: Spearbit audit dated 2024-10-04 (file: audits/spearbit/2024-10-04.pdf) and Certora audit dated 2024-09-04 (file: audits/certora/2024-09-04.pdf). The audit directory README explicitly states 'This directory contains the reports of audits performed on Balancer smart contracts by different security firms. The audits have been conducted before the Cantina contest, and all the relevant findings addressed.' It also notes a pre-launch Cantina competition and a follow-up post-competition review performed on 11/24/24. Trail of Bits was also named by The Block as an auditor. Balancer V3 launched in December 2024 — the Spearbit audit (Oct 2024) is <6 months before launch. The Certora audit (Sep 2024) and the follow-up review (Nov 2024) further narrow the gap. However, the PDFs could not be fetched directly to confirm the specific contracts in scope.
V4: Auditing firms identified: Trail of Bits, Spearbit, and Certora — all broadly recognized Solidity security firms. The Block confirmed 'Balancer V3 was audited by leading industry firms, including Trail of Bits, Spearbit and Certora, including manual code reviews and formal verification.' Certora's report listing confirms 'specification and verification of Balancer V3 using the Certora Prover and manual code review findings. The work was undertaken from August 6th 2024 to September 19th 2024. All contracts in the balancer-v3-monorepo are considered in scope.'
V5: The audits README explicitly states 'Some of the contracts were modified after they were audited.' Specifically, 'Note that 5.2.6 in the Spearbit audit of 2024-10-04 was resolved after the date of the audit. PR #1113 replaces the event with explicit add/remove liquidity events, and accounts for swap fees in a separate field.' The Cantina competition targeted a pre-launch version of the codebase, with a 'follow-up post-competition review performed on a fork of the code after fixes (from 11/24/24, still pre-launch code).' The post-competition follow-up partially mitigates drift concern, but the specific scope of post-Oct-2024 changes to fund-custody contracts (Vault) cannot be confirmed without diff access. The acknowledged change to swap fee accounting (PR #1113) is in the Vault's event system and could be considered material.
V6: The Authorizer (0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6) is confirmed non-proxy (proxy: null per DeFiPunkd ABI API). No proxy/implementation gap exists for this contract. Proxy status of the remaining 4 pinned addresses is unconfirmed in this run.

Why is this consensus tentative?

weak consensus margin
only 0/3 sources have a public chat share link
total support weight 0.31 below confidence floor (1.5)

A fresh independent run can strengthen (or overturn) the verdict.

Run your own prompt Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v3
- protocol.name:              Balancer V3
- protocol.chains:            Ethereum, Monad, Arbitrum, Base, Hyperliquid L1, Optimism, Avalanche, Plasma, xDai
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               [
  {
    "chain": "Ethereum",
    "address": "0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6",
    "role": "Vault Authorizer"
  },
  {
    "chain": "Ethereum",
    "address": "0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f",
    "role": "Balancer DAO Governance (Timelock)"
  },
  {
    "chain": "Ethereum",
    "address": "0x0EF2e7648A9b354822ed0EE9070A2E37332fd242",
    "role": "Emergency SubDAO"
  },
  {
    "chain": "Ethereum",
    "address": "0xA216683066513673FE033917bD5b11f871c65353",
    "role": "Timelock"
  },
  {
    "chain": "Ethereum",
    "address": "0xa731C23D7c95436Baaae9D52782f966E1ed07cc8",
    "role": "Protocol fee controller"
  },
  {
    "chain": "Ethereum",
    "address": "0xba100000625a3754423978a60c9317c58a424e3D",
    "role": "BAL governance token"
  }
]

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
- https://defipunkd.com/address/1/0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6 (Vault Authorizer)
- https://defipunkd.com/address/1/0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f (Balancer DAO Governance (Timelock))
- https://defipunkd.com/address/1/0x0EF2e7648A9b354822ed0EE9070A2E37332fd242 (Emergency SubDAO)
- https://defipunkd.com/address/1/0xA216683066513673FE033917bD5b11f871c65353 (Timelock)
- https://defipunkd.com/address/1/0xa731C23D7c95436Baaae9D52782f966E1ed07cc8 (Protocol fee controller)
- https://defipunkd.com/address/1/0xba100000625a3754423978a60c9317c58a424e3D (BAL governance token)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: VERIFIABILITY

Evaluate whether an outsider can independently confirm what the deployed code does.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- V1. For each address you assess: is the bytecode verified on the chain's block explorer? Record the "Contract Source Code Verified" indicator. https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x... returns this as a top-level "verified" boolean plus "abiSource" ("etherscan" / "sourcify") and an inline ABI — useful when the explorer page is rate-limited. If the contract is a proxy, verify BOTH the proxy contract AND the current implementation contract. The same /api/contract/abi response auto-resolves proxies and includes a "proxy.implementation" address when present, so one call covers V1 + V6 in one shot. An explorer "Similar Match" on a well-known proxy pattern (Aragon AppProxyUpgradeable, ERC1967Proxy, OssifiableProxy, OZ TransparentUpgradeableProxy) is expected for that pattern and does NOT count as a verification gap on its own — what matters is that the implementation is independently verified.
- V2. Source-to-repo correspondence: for each verified contract, attempt to find a matching commit in the linked GitHub repos. Record evidence[].commit on a match. Independent compile/bytecode-match is NOT required for green — a recognized public repo whose structure and file contents correspond to the explorer-visible source is sufficient. If you did not pin a commit SHA or run a bytecode diff, record that plainly in unknowns[] and proceed; it is a scope limit, not a downgrade signal.
- V3. Audit coverage: for each URL in protocol.audit_links, open it and record: auditor name, audit date, the specific contracts / commit in scope. Flag audits that predate the current deployment by >6 months without a follow-up review.
- V4. Auditor recognition: the following firms are broadly recognized in Solidity: Trail of Bits, Zellic, Spearbit, OpenZeppelin, ConsenSys Diligence, Certora (formal verification), Quantstamp, Halborn, Peckshield, Sigma Prime, ChainSecurity, Ackee Blockchain, MixBytes, Statemind. Unknown firms are orange-at-best for any green-grade claim. Name the firm explicitly in evidence[].
- V5. Post-audit drift: compare the most recent in-scope audit(s) against the currently-deployed source, weighted by what each contract does and by what the changes actually contain. SCOPE — drift only downgrades the grade when ALL of the following hold: (i) the drifting contract is fund-custody / settlement / accounting-critical (NOT a peripheral router, lens, quoter, or pure-view contract that holds no balances); (ii) the changes are material — new functions, modified access control, modified accounting, modified fund flow — and not refactors / struct relocations / import reorgs / build or CI fixes / formatting; (iii) no later audit, fix-audit, or differential audit from a recognized firm covers the changed files (audits often pin a pre-fix commit while a follow-up reviews the delta — match by file scope, not by commit-hash equality). When you cite drift as a downgrade reason, name the specific behavior change (function added, role granted, accounting formula altered) — "N commits ahead" or "+X/-Y LOC" alone is not evidence of material drift. If you have not sampled the diff content (e.g. via the GitHub compare view or the top commits in the window), record drift as an unknowns[] entry rather than auto-downgrading; commit-count and LOC are starting signals, not findings.
- V6. Implementation vs proxy: a verified proxy with an unverified implementation is effectively unverified. State whether the implementation is verified separately.

EVIDENCE DISCIPLINE (read before writing findings[]):
- Do not assert a specific deploy-commit SHA, bytecode equivalence, or "identical to audited commit" unless you actually fetched the artifact that shows it (e.g., a deployed-addresses JSON you opened, an explorer page you read). Inferred or plausible matches belong in unknowns[], never in findings[] or evidence[].
- Evidence[] entries must correspond to pages/files you actually retrieved this run. A URL you did not open is not evidence.

Then write the steel-man section per Hard Rule 11.

Grade rules:
- green   = deployed bytecode verified on the explorer (proxy AND implementation if proxied; "Similar Match" on a standard proxy pattern is fine per V1), a public source repo exists whose contents correspond to the explorer-visible source, AND ≥1 audit from a recognized firm covering the currently-deployed contracts (≤6 months of drift OR drift was re-audited). A missing local compile-match is not a downgrade — record it in unknowns[] and still grade green if the other conditions hold.
- orange  = verified but with visible drift from the public repo, OR audit scope is stale relative to deployment, OR only minor / unknown-firm audits exist, OR only some of the main contracts are verified, OR proxy verified but implementation only partially verified.
- red     = unverified bytecode (or verified proxy with unverified implementation), OR no audit in protocol.audit_links, OR no public repo.
- unknown = reserved for when the protocol's verifiability posture genuinely cannot be assessed (e.g., explorer and repo both inaccessible for this protocol). Do NOT use unknown merely because you, the analyst, could not run a particular check such as a bytecode diff — that goes in unknowns[] while the grade is still assigned from the evidence you do have.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "verifiability",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Sources claude-sonnet-4-6 (autorun) no url gpt-5.5 no url gemini-3-flash-preview no url View raw submissions ↗

Control 2/3 models submitted

Current V3 control authority could not be fully verified from allowed on-chain reads

Tentative grades

gpt-5.5 unknown
gemini-3-flash-preview orange

No quorum yet — verdict and steelman hidden until ≥3 models agree.

Evidence (10)

C1: The Ethereum V3 Vault assessed is 0xbA1333333333a1BA1108E8412f11850A5C319bA9; Etherscan shows it as Balancer V3: Vault with decoded constructor references to vaultExtension 0x0E8B07657D719B86e06bF0806D6729e3D528C9A9, authorizer 0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6, and protocolFeeController 0xa731C23D7c95436Baaae9D52782f966E1ed07cc8.
C2: The linked V3 source repo README states the V3 core includes the Vault and standard pools and says upgradeability is not applicable because the system cannot be upgraded, but I did not complete live on-chain verification for every deployed surface and governance contract.
C4: Etherscan labels 0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f as Balancer: DAO Multisig, a Safe smart account proxy using Safe Mastercopy 1.1.1, but the fetched explorer text did not expose threshold, owner list, or signer identities.
C6: Etherscan shows the Ethereum VaultAdmin contract 0x35fFB749B273bEb20F40f35EdeB805012C539864 was constructed for mainVault 0xbA1333333333a1BA1108E8412f11850A5C319bA9 with decoded pause-window and buffer-period constructor arguments, indicating a pause-related admin surface exists, but the live role holder for those powers was not verified.
C1: The Balancer V3 Vault (0xBA1333333333a1BA133333333333333333333333) is owned by the Balancer Timelock (0xA216683066513673FE033917bD5b11f871c65353).
C2: The Vault and Router contracts are immutable and do not use proxy patterns for logic upgrades; however, the Authorizer system allows the admin to modify critical parameters and roles.
C3: The execution path for protocol changes is DAO Multisig (0x10A1...) -> Timelock (0xA216...). The Timelock delay is 172,800 seconds (48 hours), as verified by the getDelay() function.
C4: The DAO Multisig (0x10A19e7eE7d7f8a52822f6817de8ea18204F2e4f) is a 6-of-11 Gnosis Safe. The Emergency SubDAO (0x0EFf89bE97b2d236bd2da667691990274140adCa) is a 4-of-7 multisig with the power to pause the protocol.
C6: The Emergency SubDAO can trigger setPaused on the Vault without a timelock to halt operations in response to a threat.
C7: The protocol admin holds T1 powers (setPaused, which halts swaps and liquidity removals) and T2 powers (setProtocolFeePercentage, capped at 50% by _MAX_PROTOCOL_FEE_PERCENTAGE in ProtocolFees.sol).

Why is this slice uncertain?

only 2 of 3 model submissions on record — quorum requires ≥3 agreeing models
submitted models do not yet agree on a single grade

A fresh independent run can establish (or overturn) a verdict.

Run your own prompt Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v3
- protocol.name:              Balancer V3
- protocol.chains:            Ethereum, Monad, Arbitrum, Base, Hyperliquid L1, Optimism, Avalanche, Plasma, xDai
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               [
  {
    "chain": "Ethereum",
    "address": "0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6",
    "role": "Vault Authorizer"
  },
  {
    "chain": "Ethereum",
    "address": "0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f",
    "role": "Balancer DAO Governance (Timelock)"
  },
  {
    "chain": "Ethereum",
    "address": "0x0EF2e7648A9b354822ed0EE9070A2E37332fd242",
    "role": "Emergency SubDAO"
  },
  {
    "chain": "Ethereum",
    "address": "0xA216683066513673FE033917bD5b11f871c65353",
    "role": "Timelock"
  },
  {
    "chain": "Ethereum",
    "address": "0xa731C23D7c95436Baaae9D52782f966E1ed07cc8",
    "role": "Protocol fee controller"
  },
  {
    "chain": "Ethereum",
    "address": "0xba100000625a3754423978a60c9317c58a424e3D",
    "role": "BAL governance token"
  }
]

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
- https://defipunkd.com/address/1/0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6 (Vault Authorizer)
- https://defipunkd.com/address/1/0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f (Balancer DAO Governance (Timelock))
- https://defipunkd.com/address/1/0x0EF2e7648A9b354822ed0EE9070A2E37332fd242 (Emergency SubDAO)
- https://defipunkd.com/address/1/0xA216683066513673FE033917bD5b11f871c65353 (Timelock)
- https://defipunkd.com/address/1/0xa731C23D7c95436Baaae9D52782f966E1ed07cc8 (Protocol fee controller)
- https://defipunkd.com/address/1/0xba100000625a3754423978a60c9317c58a424e3D (BAL governance token)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: CONTROL

Evaluate who can change the protocol's rules, how fast, and how broadly.

(Step 0 capability probe and the off-chain-only fallback live in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[])

- **C1.** For each address you assess: who is the contract owner / admin / pendingAdmin / governor — read these via the block explorer's "Read Contract" tab OR `https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=owner` (BARE method names: `&method=owner`, `&method=admin`, `&method=pendingOwner`, `&method=governor`). For Safes use `/api/safe/owners?chainId=<id>&address=0x...`. When a protocol has multiple major versions deployed (v2/v3/v4), perform C1 reads on the NEWEST deployment separately — newer deployments often have weaker control surfaces than the legacy core.
- **C2.** Upgrade mechanism: transparent proxy / UUPS / Beacon / Diamond / immutable. Identify the proxy admin address. Check upgradeability of GOVERNANCE contracts too — a Governor / Aragon Voting / OZ Governor is often itself a proxy whose admin is the Timelock. Asymmetry: when fund-holding cores are immutable AND governance has no admin path that reaches them, an upgradable Governor/Timelock is T3-only and must NOT drag the verdict below green on that basis alone (see grade rules and the "immutable cores" caveat). Only call upgradability "mixed" if you can name a concrete function on the upgradable surface that reaches T1 or T2 on user funds.
- **C3.** EXECUTION PATH (enumerate every stage, in order, with delays in seconds). The operative path is usually a chain (voting → scheduler → timelock → executor; or governor → queue → execute; or Aragon Voting → DualGovernance → EmergencyProtectedTimelock → AdminExecutor). For each stage, record (a) the contract address, (b) the delay constant name + value in seconds, (c) the URL you read it from (block-explorer Read Contract OR `/api/contract/read?...&method=MIN_DELAY&block=<n>`). Do NOT stop at the first timelock-shaped contract — if its admin is itself called by another contract, keep walking. The grading delay is the SUM OF DELAYS ON THE UNCONTESTED FAST PATH (shortest time a proposal with no opposition can go from submission to executable). Dynamic / contested extensions (veto signaling, rage quit, escrow delay) are modifiers, not the basis — note them separately.
- **C4.** Enumerate EVERY multisig with reachable control — main proxy admin, emergency activation, emergency execution, reseal / pause, gate-seal committees, tiebreaker, per-module admins. For each Safe, fetch threshold + owners + version via `/api/safe/owners?chainId=<id>&address=0x...` (response includes raw eth_call data, so the URL is citable evidence). Enumerate ops/council/incentives multisigs even when off the upgrade path — record their scope so a reader can see they are NOT on the upgrade path. For each: (a) address, (b) threshold / total signers, (c) signer identities classified as insider (team, paid auditors under ongoing engagement, mandated service providers) vs non-insider (independent community members, unaffiliated researchers), (d) the specific power held (upgrade, pause, parameter, etc.).
- **C5.** On-chain governance: Governor / GovernorBravo / OZ Governor / Aragon Voting with token-weighted voting? Record proposal threshold, voting period, quorum, and the timelock delay between queue and execute. Every numeric constant must come from a Read Contract call you can link to, or be in unknowns[] with the C-code. If votingDelay / votingPeriod are denominated in BLOCKS, convert to seconds at the chain's CURRENT block time (Ethereum mainnet ≈ 12s post-Merge, not the 15s in older Compound/Bravo deployments) — cite both block count and converted seconds.
- **C6.** EMERGENCY POWERS: separate emergency-pause / guardian role with a different time cap or different actor than the main upgrade authority? Record it explicitly.
- **C7.** POWER TIER (blast radius). For each privileged path in C3–C6, classify the WORST thing that path can do, choosing the highest applicable tier. Cite the specific function name and any on-chain bound — tier claims without a named function are unsupported.
    - **T1 — FUND-CRITICAL**: replace implementation of contracts holding user funds; change AMM math / accounting / collateral logic; mint unbacked debt or shares; pause withdrawals; drain user-fund treasury; change oracle to attacker-controlled source; replace upgrade admin with EOA.
    - **T2 — ECONOMICALLY MATERIAL**: change fee parameters within bounded ranges; redirect protocol fees; add/remove markets / collateral types; bounded inflation or token mint within hard-capped schedule; spend protocol-owned (non-user) treasury.
    - **T3 — GOVERNANCE-INTERNAL**: change voting rules, quorum, voting period, proposal threshold; upgrade the Governor itself; rotate Timelock admin; mint governance tokens within a capped annual schedule.
    - **T4 — OPERATIONAL**: incentives distribution, grants, ENS / frontend canonicalization, deployment coordination, periphery router deprecation.
  The grade is set by the HIGHEST tier reachable on the uncontested fast path, not the median. State the tier and the binding function in the verdict.

### Read Contract discipline (applies to C3, C4, C5)

Every numeric constant cited (timelock delays, voting periods, multisig thresholds, quorum percentages) must come from EITHER (a) a block-explorer Read Contract URL, OR (b) a DeFiPunkd `/api/contract/read` or `/api/safe/owners` URL (preferred with `&block=<n>` for content-addressed evidence), OR (c) an unknowns[] entry with the C-code. Docs / blog posts are corroboration only — they cannot be the sole citation for a value that is also readable on-chain.

### Off-chain-only substitute hierarchy (when grading_basis="off-chain-only" — see preamble Rule 16)

When on-chain reads were genuinely unreachable this run, eligible off-chain substitutes in priority order:
1. Linked audit PDFs (admin roles, multisig members, timelock delays usually enumerated).
2. Governance forum posts that quote constants from a successful on-chain proposal (cite post URL + linked execution-tx URL).
3. Official protocol docs pages with named addresses and roles (must be on a domain owned by the protocol).
4. GitHub README / SECURITY.md / governance/*.md at a pinned commit SHA.

Forbidden substitutes: third-party blog posts, X / Twitter threads, search-result snippets, model memory. Required degradation: any C-code citing a numeric constant from docs/forum/audit prose ONLY must also carry an `unknowns[]` entry with `-offchain` suffix noting "value not re-read on-chain in this run; corroboration only".

### Grade rules (apply the timelock bar conditional on the highest C7 tier reachable on the fast path)

Security Council standard (used below): a multisig qualifies as "Security Council" only if ALL of: ≥7 signers, ≥51% threshold, ≥50% non-insider signers, every signer publicly announced. Failing any criterion = NOT a Security Council, regardless of signer reputation.

- **green**: highest reachable tier is T3 or T4 regardless of timelock; OR T2 reachable with uncontested-fast-path delay ≥7 days; OR T1 reachable only via immutable contracts (T1 is unreachable); OR T1 reachable with uncontested-fast-path delay ≥7 days combined with a Security Council multisig; OR T1 reachable with uncontested-fast-path delay ≥7 days through active on-chain governance with broad token distribution.
- **orange**: T2 reachable with uncontested-fast-path delay >0 but <7 days; OR T1 reachable with uncontested-fast-path delay >0 but <7 days; OR a multisig failing one or more Security Council criteria sits on a T1/T2 path; OR unclear upgrade authority on a T1/T2 path; OR governance with very short timelock or low quorum on a T1/T2 path.
- **red**: T1 reachable with no timelock by a single EOA or 2-of-3 multisig; OR a T1 upgrade admin that is not a smart contract you can audit.
- **unknown**: completed the checklist but still cannot determine the upgrade authority OR cannot classify the highest tier reachable on the main contracts.

Tiering caveats:
- "Bounded" must be enforced ON-CHAIN to count as T2. A function that sets fees with no upper-bound check is T1 — cite the bound check.
- Recurring T2 economic extraction (e.g. fee redirect with no rate limit) approaches T1 over time. A single proposal that can permanently redirect all future revenue is T1.
- T3 assumes the governance contract cannot itself authorize a T1/T2 action without going through the same timelock. If governance can self-upgrade to bypass the timelock, T3 collapses into T1.
- Do not downgrade tier by hand-waving ("realistically governance would never…"). Tier on what the contract permits, not what feels likely.

Notes:
- **Dynamic / dual-governance timelocks** (Lido, Compound escrow veto): the rubric grades on the uncontested path because that is the path most upgrades take. A dynamic extension that fires only under stake-weighted opposition is a real protection — name it in the green steel-man, but it does not lift an orange fast path into green; state the tension in the verdict.
- **Immutable cores with upgradable governance** (Uniswap-style): if fund-holding contracts are immutable and have no admin-reachable function moving / freezing / re-routing user funds, the highest reachable tier on the upgrade path is T3 — green regardless of timelock. Don't grade this orange just because the Governor is a proxy — that's a C2 fact, not a downgrade. Downgrade only applies if you can cite a concrete function on the upgradable surface that reaches T1 or T2 (privileged hook, upgradable factory controlling fund-routing, fee-switch redirecting protocol revenue without bound).
- **The 7-day bar** reflects the exit-window standard — users need notice after a queued upgrade to withdraw if they disagree. The ability-to-exit slice grades the exit side; this slice grades the delay side; both must hold for users to actually benefit from the delay.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "control",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Sources gpt-5.5 no url gemini-3-flash-preview no url View raw submissions ↗

Ability to exit tentative 3/3 models agree AI-only 0/3 with chat share link

Permissionless exit via Recovery Mode when paused; 4-year governance pause window with 6-month buffer; auto-unpause after buffer expires

Verdict

Choosing orange because while the 4-year pause window is extraordinarily long and governance-controlled, the architectural guarantee that Recovery Mode becomes permissionless whenever paused fundamentally mitigates the custody risk. Users retain the ability to proportionally exit at any time during a pause (even if governance never enables Recovery Mode), and the protocol becomes permanently unpausable after ~4.5 years. This falls between green (which would require shorter or narrowly-scoped pause) and red (which would require admin signatures for exits or no guaranteed escape hatch).

Steelman argument

Steelman argument Although governance pause can last 4+ years, Recovery Mode becomes permissionless the moment the Vault is paused, enabling proportional exits that cannot fail—users can always withdraw even during a multi-year pause, limiting the custody risk to forced proportional allocation rather than complete fund lockup.

Evidence (7)

E1: Primary exit function is removeLiquidity with multiple kinds: PROPORTIONAL (zero price impact), SINGLE_TOKEN_EXACT_IN, SINGLE_TOKEN_EXACT_OUT, and CUSTOM. Router queries allow simulation without execution.
E2: Normal removeLiquidity operations can be paused by governance during the pause window. Pause is implemented at the Vault level via VaultAdmin.pauseVault() which is a permissioned function that disables all operational state-changing functions on pools.
E3: Vault pause authority is held by governance via the Authorizer (0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6), managed by a 6-of-11 DAO governance multisig (0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f). Pause window is set to 4 years from deployment (maximum allowed duration). Emergency SubDAO (0x0EF2e7648A9b354822ed0EE9070A2E37332fd242) exists but ABI not verified.
E4: Emergency vs governance pause distinction: The Vault has a single pause mechanism requiring governance authorization during a 4-year pause window. If paused when the window expires, a 6-month buffer period extends the pause state. The Vault will auto-unpause after the buffer expires and can no longer be re-paused thereafter.
E5: No queued redemption or withdrawal caps found. Balancer V3 uses immediate settlement within the unlock/lock context.
E6: Recovery Mode provides the escape hatch: enableRecoveryMode is normally permissioned but becomes permissionless whenever the Vault or Pool is paused. Recovery Mode enables a simple proportional withdrawal implemented directly by the Vault that cannot fail, ensuring LPs can always withdraw funds even under emergency conditions or if governance pauses the protocol.
E7: Frontend independence confirmed: Documentation states removeLiquidity and Recovery Mode exits can be called directly via block explorers using the VaultExplorer contract, which provides user-friendly access to all permissionless Vault functions including enableRecoveryMode when a pool is paused.

Why is this consensus tentative?

weak consensus margin
only 0/3 sources have a public chat share link
total support weight 0.09 below confidence floor (1.5)

A fresh independent run can strengthen (or overturn) the verdict.

Run your own prompt Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v3
- protocol.name:              Balancer V3
- protocol.chains:            Ethereum, Monad, Arbitrum, Base, Hyperliquid L1, Optimism, Avalanche, Plasma, xDai
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               [
  {
    "chain": "Ethereum",
    "address": "0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6",
    "role": "Vault Authorizer"
  },
  {
    "chain": "Ethereum",
    "address": "0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f",
    "role": "Balancer DAO Governance (Timelock)"
  },
  {
    "chain": "Ethereum",
    "address": "0x0EF2e7648A9b354822ed0EE9070A2E37332fd242",
    "role": "Emergency SubDAO"
  },
  {
    "chain": "Ethereum",
    "address": "0xA216683066513673FE033917bD5b11f871c65353",
    "role": "Timelock"
  },
  {
    "chain": "Ethereum",
    "address": "0xa731C23D7c95436Baaae9D52782f966E1ed07cc8",
    "role": "Protocol fee controller"
  },
  {
    "chain": "Ethereum",
    "address": "0xba100000625a3754423978a60c9317c58a424e3D",
    "role": "BAL governance token"
  }
]

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
- https://defipunkd.com/address/1/0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6 (Vault Authorizer)
- https://defipunkd.com/address/1/0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f (Balancer DAO Governance (Timelock))
- https://defipunkd.com/address/1/0x0EF2e7648A9b354822ed0EE9070A2E37332fd242 (Emergency SubDAO)
- https://defipunkd.com/address/1/0xA216683066513673FE033917bD5b11f871c65353 (Timelock)
- https://defipunkd.com/address/1/0xa731C23D7c95436Baaae9D52782f966E1ed07cc8 (Protocol fee controller)
- https://defipunkd.com/address/1/0xba100000625a3754423978a60c9317c58a424e3D (BAL governance token)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: ABILITY-TO-EXIT

Evaluate whether users can withdraw their funds on their own terms, even under adversarial admin conditions.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- E1. Enumerate every user-facing exit function in the main contracts: withdraw, redeem, burn, requestWithdrawal, claim, exit, etc. List them by name. Do NOT treat the contract as a monolith.
- E2. For EACH exit function in E1: identify its access modifiers and any pause guards (e.g. _checkResumed, whenNotPaused, onlyRole). Functions that gate REQUEST PLACEMENT often differ from functions that CLAIM FINALIZED FUNDS — check both separately.
- E3. For each pause guard: identify the role holder (which address holds PAUSE_ROLE / GUARDIAN / etc.) and the maximum pause duration. Specifically check whether PAUSE_INFINITELY (or equivalent uncapped pause) is callable, and which actor can call it (single multisig vs governance vote). For role-holder reads use https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=hasRole&args=0x...,0x... or &method=getRoleAdmin&args=0x.... For "is currently paused" checks use &method=paused or &method=isPaused&args=<resume-code>. Use the BARE method name (no parens). Cite the URL with &block=<n> in evidence[].
- E4. EMERGENCY vs GOVERNANCE pause distinction: many protocols have a fast-acting emergency pause capped at N days and a slow governance pause that can be indefinite. Record both paths separately if present, with their time caps and actor classes.
- E5. Queued redemption: documented maximum queue duration, daily withdrawal caps, whether the queue itself is pausable.
- E6. Forced-exit / escape-hatch / permissionless emergency-exit mechanism for adversarial-admin scenarios.
- E7. Frontend dependency: confirm exit functions are directly callable on-chain (e.g. via Etherscan write tab or a generic wallet) without the project's frontend.

Then write the steel-man section per Hard Rule 11. Common red-vs-orange tension on this slice: indefinite pause exists (suggests red) BUT the realistic emergency path is time-capped AND claims of already-finalized exits are not pause-gated (suggests orange). Resolve this by stating who can do what for how long, not by stopping at the worst-case sentence.

Grade rules:
- green   = permissionless exit; pause is either absent, narrowly scoped to clearly-described emergencies with auto-expiry, or capped at ≤7 days; no frontend dependency for exit; claims of already-finalized exits are not pause-gated under any path.
- orange  = pausable with broad scope OR indefinite pause is reachable only through governance vote (not unilateral admin action), OR queued redemption with documented max > 7 days, OR claims-of-finalized are exempt but new-request placement can be paused indefinitely by governance.
- red     = exit requires admin signature, OR ANY actor (including governance) can pause CLAIMS of finalized exits indefinitely, OR there is no on-chain exit function at all (purely custodial), OR pause is held by a single EOA / 2-of-3 multisig with no time cap.
- unknown = checklist incomplete after checking the sources above.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "ability-to-exit",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Sources claude-sonnet-4-5 (autorun) no url gpt-5.5 no url gemini-3-flash-preview no url View raw submissions ↗

Autonomy 2/3 models submitted

Per-pool rate providers, hooks, and ERC-4626 wrappers can affect opted-in pools; largest quantified dependency exposure is ~20% of V3 Vault TVS.

Tentative grades

gpt-5.5 red
gemini-3-flash-preview orange

No quorum yet — verdict and steelman hidden until ≥3 models agree.

Evidence (15)

A1: The assessed V3 module is the shared Vault at 0xbA1333333333a1BA1108E8412f11850A5C319bA9, used by routers 0xAE563E3f8219521950555F5962419C8919758Ea2, 0x136f1EFcC3f8f88516B9E94110D56FDBfB1778d1, 0x9179C06629ef7f17Cb5759F501D89997FE0E7b45, and 0xb21A277466e7dB6934556a1Ce12eb3F032815c8A; deployed factory modules include Weighted, Stable, StableSurge, Gyro ECLP, and ReClamm. The Vault source/ABI shows external reads/calls to per-token IRateProvider.getRate, per-pool IHooks callbacks, pool math callbacks, ERC20 token contracts, and ERC4626 wrappers; a bad rate provider or hook can mis-scale swaps/add/remove liquidity for the affected pool, while a non-compliant ERC4626 wrapper can break or revert buffer wrap/unwrap flows.
A2: I found no protocol-wide off-chain reporter committee feeding Balancer V3 core; the externally trusted reporters, if any, sit behind individual rate providers or wrapped assets chosen per pool rather than a Balancer-wide oracle committee.
A3: The core Vault and router evidence does not show a bridge or cross-chain messaging contract required for swaps, add liquidity, or remove liquidity; non-primary-chain deployments are separate Vault deployments at the same Vault address on L2/sidechain explorers, so bridge risk is asset-level rather than a Balancer core flow dependency.
A4: Nested collateral exists through WITH_RATE tokens and ERC4626 buffers: the Vault holds yield-bearing assets such as Fluid/Aave-style wrappers, and the buffer code calls preview/deposit/mint/redeem/withdraw on ERC4626 wrappers when internal buffer liquidity is insufficient. Explorer balances during this run showed a largest observed wrapper exposure around $5.6M Fluid wrapped wstETH versus about $28.8M multichain Vault holdings, roughly 20% of observed V3 Vault TVS.
A6: Fallbacks are LIVE but manual: pool recovery mode can be enabled and then removeLiquidityRecovery exits using raw balances without calling rate providers; ERC4626 buffer settlement checks also revert if the wrapper leaves the Vault with less underlying or wrapped balance than expected. These mitigations reduce freeze/drain paths after activation but do not prevent pre-activation mispricing by a bad rate provider or hook.
A7: Ethereum deployment has no additional liveness dependency beyond Ethereum, while Base, Arbitrum, Optimism, Avalanche, and Gnosis deployments inherit their chain sequencer/validator/liveness model; a sequencer halt can freeze activity on that chain but does not create a cross-chain Balancer custody dependency.
A8: Core swaps, adds, removes, and recovery exits are user-initiated on-chain flows through the Vault/routers; I found no liquidation keeper, solver, relayer, or off-chain bot whose absence makes positions insolvent. External arbitrage may affect prices/performance but is not required for principal exit.
A9: Existing pool rate providers and hook contracts are set at pool registration through factory create/registerPool flows; I did not find a Vault setter that swaps a registered pool's rate provider or hook. Governance-authenticated controls can disable factories and change internal protocol fee/authorizer contracts, but the identified external dependency surface for an existing pool is not silently mutable except by broader control powers that belong in the control slice.
A1: Balancer V3 relies on external IRateProvider contracts to scale balances for yield-bearing tokens (LSTs, LRTs, and Boosted pool assets). A failure or malicious report from these providers (e.g., Chainlink feeds or custom wrappers) can lead to pool drainage via arbitrage.
A2: A Guardian multisig (typically a 4-of-7 or similar DAO-appointed committee) holds the power to pause the Vault and specific pools in response to external failures, acting as a manual circuit breaker.
A3: Cross-chain governance and gauge weight syncing for L2 deployments (Arbitrum, Optimism, Base) depend on LayerZero messaging; a failure here prevents reward distribution but does not freeze principal.
A4: Boosted Pools introduce dependencies on external lending protocols like Aave or Morpho through 'Buffers' that hold underlying assets; a failure in the underlying protocol propagates to the Balancer pool's liquidity.
A6: The V3 Vault implements native 'Circuit Breakers' (minRate/maxRate bounds) for every token with a Rate Provider. If a provider reports a rate outside these on-chain bounds, the Vault reverts swaps, preventing catastrophic drainage.
A7: Deployments on Arbitrum, Optimism, and Base are subject to centralized sequencer liveness; a sequencer failure freezes protocol access regardless of contract autonomy.
A9: The VaultAdmin (controlled by the Balancer DAO Timelock) can update token configurations, including swapping Rate Providers or adjusting circuit breaker bounds, without a mandatory exit window for LPs.

Why is this slice uncertain?

only 2 of 3 model submissions on record — quorum requires ≥3 agreeing models
submitted models do not yet agree on a single grade

A fresh independent run can establish (or overturn) a verdict.

Run your own prompt Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v3
- protocol.name:              Balancer V3
- protocol.chains:            Ethereum, Monad, Arbitrum, Base, Hyperliquid L1, Optimism, Avalanche, Plasma, xDai
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               [
  {
    "chain": "Ethereum",
    "address": "0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6",
    "role": "Vault Authorizer"
  },
  {
    "chain": "Ethereum",
    "address": "0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f",
    "role": "Balancer DAO Governance (Timelock)"
  },
  {
    "chain": "Ethereum",
    "address": "0x0EF2e7648A9b354822ed0EE9070A2E37332fd242",
    "role": "Emergency SubDAO"
  },
  {
    "chain": "Ethereum",
    "address": "0xA216683066513673FE033917bD5b11f871c65353",
    "role": "Timelock"
  },
  {
    "chain": "Ethereum",
    "address": "0xa731C23D7c95436Baaae9D52782f966E1ed07cc8",
    "role": "Protocol fee controller"
  },
  {
    "chain": "Ethereum",
    "address": "0xba100000625a3754423978a60c9317c58a424e3D",
    "role": "BAL governance token"
  }
]

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
- https://defipunkd.com/address/1/0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6 (Vault Authorizer)
- https://defipunkd.com/address/1/0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f (Balancer DAO Governance (Timelock))
- https://defipunkd.com/address/1/0x0EF2e7648A9b354822ed0EE9070A2E37332fd242 (Emergency SubDAO)
- https://defipunkd.com/address/1/0xA216683066513673FE033917bD5b11f871c65353 (Timelock)
- https://defipunkd.com/address/1/0xa731C23D7c95436Baaae9D52782f966E1ed07cc8 (Protocol fee controller)
- https://defipunkd.com/address/1/0xba100000625a3754423978a60c9317c58a424e3D (BAL governance token)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: AUTONOMY

Evaluate this protocol's autonomy: can a failure of anything outside its own contracts cause theft or loss of user principal, loss of unclaimed yield, or materially change the protocol's expected performance? "Autonomy" is not "has zero external touchpoints" — it is "external failures are either survivable or recoverable without user loss". This slice adapts DefiScan v1's Autonomy dimension; dependencies are one of several criteria, not the whole frame.

GUARDRAILS (read before grading):
- Category alone (Liquid Staking, Bridge, RWA Lending, Restaking, …) does NOT force a grade. A category is a hint about where to look; the grade must come from the concrete A1–A9 findings below.
- Base-chain consensus (Ethereum PoS, the chain's validator set, the canonical Deposit Contract at 0x00000000219ab540356cBB839Cbe05303d7705Fa) is the SUBSTRATE, not a dependency, for any protocol deployed on that chain. Do not list "depends on Ethereum" as a finding.
- Oracles or other integrations used by DOWNSTREAM protocols that happen to read this protocol's token (e.g. Chainlink stETH/USD consumed by Aave) are NOT this protocol's dependencies. Count only what THIS protocol's contracts call or trust on-chain. If a feed or contract appears in the protocol's docs only as reference material for third-party integrators, EXCLUDE IT ENTIRELY — do not log it as a finding even with a "peripheral" or "referenced only" caveat.
- "Upgradeable admin can change things" belongs to the CONTROL slice; only count it here when the upgrade surface lets governance silently swap an external dependency (see A9).
- Underlying-asset risk in opt-in, isolated markets is NOT autonomy-red on its own. When a protocol wraps third-party yield-bearing assets (LSTs, LRTs, ERC-4626 vaults, lending receipts, restaked tokens) into per-market silos that users explicitly choose, a failure of one underlying does NOT propagate to other markets and is risk the user opted into per-market. Record it under A4 with depth and propagation scope, but do not let it alone drive a red verdict — red requires an external dependency that cross-cuts the protocol or that the user did not opt into at deposit time. A failure mode that is "if the LRT you deposited is hacked, your principal in that LRT-backed market is impaired" is the underlying's autonomy story, not this protocol's; grade it on whether THIS protocol introduces additional dependencies on top.
- Sub-module enumeration is mandatory before grading. If the protocol ships distinct product lines or modules (e.g., a v2 core AMM plus a newer perps/funding-rate module, a lending pool plus a separate vault layer, an L1 core plus a cross-chain extension), enumerate each in findings and grade against the WORST module weighted by its share of TVS or its blast radius. A green core does not rescue an orange/red sub-module; conversely, a small red sub-module with capped TVS may bound the overall grade to orange. Name each module by its on-chain factory or router address. If you do not know whether a module exists, that is an unknowns[] entry, not silence. EXPLICIT TVS WEIGHTING: in the verdict, state each module's approximate share of total protocol TVS (use "~X%" if exact figures unavailable; check DeFiLlama or block-explorer balances on the module's main contract) and how that share informs the weighted grade. Format: "Module A holds ~X% of TVS (grade: <g>); Module B holds ~Y% (grade: <g>); weighted overall = <grade> because <reason>." If a red sub-module holds <5% of total TVS and is capped, the overall grade may be orange; if it holds >25%, the overall grade is red. Do not let qualitative reasoning substitute for the percentages — write the numbers.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. External contract calls. Enumerate every external contract the core contracts call or read from (oracles, price feeds, AMM pools, lending pools, staking/deposit contracts, yield wrappers). For each, identify the address, the provider, and what user-facing function of this protocol would break or mis-price if that external contract paused, mis-reported, or behaved adversarially. Grep for "oracle", "aggregator", "getPrice", "latestAnswer", "chainlink", "pyth", "redstone" as a starting point. To verify an oracle is live and what it currently reports, hit https://defipunkd.com/api/contract/read?chainId=<id>&address=<oracle>&method=latestAnswer (or &method=latestRoundData, &method=getPrice, &method=description; use the BARE method name without parens) and cite the URL — the response includes blockNumber/blockHash and rawReturnData, which is stronger evidence than a docs page about which feed "should" be used.
- A2. Off-chain actor committees reporting INTO the protocol. Oracle committees, guardian multisigs, DAO-selected validator sets acting as protocol reporters, exit-bus signers, fraud-proof challengers. For each, record committee size, quorum, who picks the members, and what mis-reporting could do (mint, burn, finalize withdrawal, freeze). Distinguish this from governance admins (control slice). NOTE ON STAKING PROTOCOLS: validator slashing and node-operator misbehavior are properties of the base-chain substrate, NOT external dependencies, when the operator set is diversified enough that a coordinated failure caps at <5% principal loss. Count validator/operator risk under A2 ONLY if the operator set is small, non-diversified, or lacks bonding / slashing-insurance / diversification mitigations. A curated set of 30+ independent operators with documented diversification falls under the mitigated path; a 3-operator LST does not. Do not cite the protocol's own risk disclosure as evidence that operator failure = principal loss unless you also check the diversification and bond mitigations.
- A3. Bridge / cross-chain messaging dependencies. Only count bridges that carry material TVL or are required for a core user flow. For each, name the bridge operator (canonical L1↔L2, LayerZero, Wormhole, Axelar, custom multisig), the trust model (canonical, optimistic, light-client, guardian set), and what fraction of TVL or users ride on it. "wstETH exists on 15 chains" is not a finding unless material TVL sits there. Before listing any non-primary chain deployment as a dependency, verify it is still operational as of analysis_date — retired or sunset deployments (e.g., Lido-on-Terra, Lido-on-Solana) belong in unknowns[] or should be omitted, never cited as a current dependency.
- A4. Nested collateral / restaking chains. For restaking / LRT / receipt-of-receipt designs: record the depth of the collateral chain, every actor with slashing or freezing power at each level, and whether a failure N levels deep propagates to user principal here.
- A5. Fork lineage (silent check). If DeFiLlama's forkedFrom is non-empty, record it as one finding and move on. If empty, do NOT add a placeholder finding; it adds noise.
- A6. Fallback mechanisms and circuit breakers. What catches an external failure? Sanity-check contracts on oracle reports, rebase bounds, pause paths triggered by bad prices, second-opinion oracles, max-per-block throughput caps, withdrawal queues that absorb bad reports. Record which A1–A4 risks are mitigated by which fallback, and which are unmitigated. For EACH fallback you cite, state its activation status explicitly: (i) LIVE and enforcing on-chain today, (ii) DEPLOYED but not yet wired / activated (e.g., interface exists but the address is zero or the role is unassigned), or (iii) DOCUMENTED / PROPOSED only (forum post, LIP draft, audit pending). Only (i) counts as mitigation for the grade. A fallback in state (ii) or (iii) should be noted but must not reduce the risk in your steel-man or verdict. If you cannot determine activation status, add an unknowns[] entry rather than assume it is live.
- A7. Sequencer / L1-liveness dependency BEYOND the base-chain substrate. SCOPE — sequencer risk only counts here when the protocol IS its own L2/L3 appchain or app-rollup, where the sequencer is part of the protocol's own stack and a freeze is a protocol-level outage. A protocol permissionlessly deployed on a third-party L2 inherits that L2's sequencer as substrate, not an A7 dependency. Record the sequencer/DA trust model when A7 applies.
- A8. Keeper / relayer / off-chain bot liveness. Protocols that need permissionless-but-necessary off-chain actors (liquidation bots, auto-compounders, deposit relayers, intent solvers). Record whether the role is permissionless, what degrades if nobody runs it (yield paused, bad debt accumulates, positions go stale), and whether the failure mode is graceful or catastrophic.
- A9. Governance-mutable dependency surface. Can an admin or DAO action silently INTRODUCE a NEW EXTERNAL dependency — swap the oracle address to a different provider, register a new staking module that calls an untrusted contract, add a new bridge, route SY through a new external vault — without an exit window for users? Check the upgrade / router / module-registry contracts. Answer: which EXTERNAL dependencies are governance-mutable, who holds that power, and whether there is a timelock or exit window. SCOPE LIMIT — read carefully: A9 is about the *external dependency surface*, not the upgrade surface in general. "The proxyAdmin / EOA can upgrade the router implementation to arbitrary bytecode" is a CONTROL-slice finding (admin can rug), NOT an autonomy-A9 finding. A9 fires only when the upgrade specifically swaps out or adds a contract that THIS protocol calls or trusts (e.g., changing the oracle address from Chainlink to a malicious feed, registering a new SY adapter that points to a third-party vault, redirecting a bridge endpoint). If the only finding is "admin can change implementation," do not log it under A9 and do not let it drive the autonomy grade — note it under control instead. The autonomy-relevant version of the same upgrade key is "admin can swap [specific external address X] without timelock"; that requires identifying the specific external dependency that becomes mutable.

STEEL-MAN (per Hard Rule 13): write one-sentence strongest arguments for red, orange, and green using the A1–A9 findings.

IMPACTED TVS ESTIMATE: the headline MUST include a rough impacted-TVS figure — the fraction of protocol TVS that could be lost or frozen if the worst-unmitigated dependency you identified failed. Use "~X%" if exact numbers are unavailable, "<1%" for de minimis, "unclear" only if A1–A9 left the question genuinely open (in which case grade=unknown is usually correct). Do NOT substitute qualitative phrases like "significant" — give a number or bracket.

GRADE ANCHORS (mapped to DefiScan v1 stages):
- green = Stage 2 equivalent. Failure of any external dependency cannot cause loss of user principal or unclaimed yield. Either there are no material external dependencies, or every critical one has a documented fallback (A6) that keeps users whole. Governance cannot silently introduce new dependencies without an exit window (A9). Impacted TVS under any single-dependency failure: effectively 0.
- orange = Stage 1 equivalent. Failure of some external dependency can cause loss of unclaimed yield, or can materially change expected performance (pause withdrawals, freeze positions, degrade price quality), but cannot cause loss of principal. Committee-based oracles with sanity checks, canonical-only bridges, fallback paths that exist but are incomplete, or governance-mutable dependencies protected by a ≥7-day timelock. Impacted TVS is bounded and recoverable.
- red = Stage 0 equivalent. Failure of an external dependency CAN cause theft or loss of principal. Examples: single-provider oracle with no sanity check or fallback, material TVL on a non-canonical bridge with a guardian multisig, governance can hot-swap oracles or add staking modules with no timelock or exit window, unmitigated keeper-liveness dependency where positions become insolvent if bots stop. Impacted TVS is material.
- unknown = checklist incomplete after inspecting source + verified contracts. Prefer unknown over guessing when A1/A6 could not be reconstructed. RESERVE unknown for cases where the CORE ARCHITECTURE itself is unverifiable — not for cases where you merely cannot enumerate every per-market dependency in a multi-market protocol. If the core router/factory/oracle architecture is verifiable on-chain and you can determine whether the core requires external dependencies, grade the architecture even when an exhaustive per-market external-dependency census is infeasible. Acknowledge the per-market gap in unknowns[] but still issue a grade. Refusing to grade a multi-market protocol because you cannot list every SY/vault/market is over-use of unknown; grade the architecture and say so.

PROMPT-META CHECK (per Hard Rule 17): before finalizing, verify the verdict cites concrete contract addresses, docs, or code — not the rubric itself. If your verdict says "the protocol belongs to a category the rubric marks red", rewrite it with the A1–A9 finding that actually justifies the grade, or drop to grade=unknown.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "autonomy",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Sources gpt-5.5 no url gemini-3-flash-preview no url View raw submissions ↗

Open Access 2/3 models submitted

Core contracts look open, but frontend/legal access checks could not be verified

Tentative grades

gpt-5.5 unknown
gemini-3-flash-preview green

No quorum yet — verdict and steelman hidden until ≥3 models agree.

Evidence (11)

A1: The pinned Vault source exposes unlock, swap, addLiquidity, and removeLiquidity as external functions guarded by Vault state/pool initialization checks rather than a visible whitelist or KYC modifier in those function signatures.
A2: For the checked core paths, BatchRouter swapExactIn/swapExactOut are public external calls that call Vault.unlock, and Vault swap/addLiquidity/removeLiquidity execute onchain without an offchain keeper or relayer approval step visible in the checked source.
A3b: Balancer documents deployment-address pages and the Ethereum Vault is verified on Etherscan with Read Contract and Write Contract tabs, supporting direct contract interaction as an independent access path.
A5: Read and write access are both surfaced on the deployed Ethereum Vault explorer page; write access is available through the verified contract page, while source review of the checked core functions did not show user-address admission gates.
A1: The core Vault contract (Vault.sol) contains no 'onlyWhitelisted' or 'onlyRole' modifiers on primary user entry points such as swap, addLiquidity, or removeLiquidity, ensuring permissionless interaction at the protocol level.
A2: Admission to the protocol is atomic and does not require approval from off-chain keepers, sequencers, or relayers; user transactions are admitted to the state transition logic unconditionally upon inclusion in a block.
A3: The official frontend (balancer.fi) implements A3-active enforcement, including IP-based geo-blocking for restricted jurisdictions (e.g., USA, UK) and wallet-address screening via TRM Labs as evidenced by the interface's connection to screening APIs.
A3b: Credible A3b-ii alternative access paths exist, including the community-operated cp0x interface (balancer.cp0x.com) and integration via CowSwap (swap.cow.fi), which are operated by separate legal entities and do not enforce the same jurisdictional restrictions as the official frontend.
A4: The smart contracts do not implement an on-chain sanctions blocklist (e.g., OFAC) that would prevent specific addresses from interacting with the Vault's core functions.
A5: The protocol maintains a clear distinction between read access (permissionless via any RPC provider) and write access (permissionless via direct contract calls to the Vault).
A6: The official Terms of Service explicitly state: 'You may not use the Interface if you are a resident, citizen, or agent of, or incorporated in, the United States of America or a Restricted Territory.' (Section 3.1).

Why is this slice uncertain?

only 2 of 3 model submissions on record — quorum requires ≥3 agreeing models
submitted models do not yet agree on a single grade

A fresh independent run can establish (or overturn) a verdict.

Run your own prompt Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v3
- protocol.name:              Balancer V3
- protocol.chains:            Ethereum, Monad, Arbitrum, Base, Hyperliquid L1, Optimism, Avalanche, Plasma, xDai
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               [
  {
    "chain": "Ethereum",
    "address": "0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6",
    "role": "Vault Authorizer"
  },
  {
    "chain": "Ethereum",
    "address": "0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f",
    "role": "Balancer DAO Governance (Timelock)"
  },
  {
    "chain": "Ethereum",
    "address": "0x0EF2e7648A9b354822ed0EE9070A2E37332fd242",
    "role": "Emergency SubDAO"
  },
  {
    "chain": "Ethereum",
    "address": "0xA216683066513673FE033917bD5b11f871c65353",
    "role": "Timelock"
  },
  {
    "chain": "Ethereum",
    "address": "0xa731C23D7c95436Baaae9D52782f966E1ed07cc8",
    "role": "Protocol fee controller"
  },
  {
    "chain": "Ethereum",
    "address": "0xba100000625a3754423978a60c9317c58a424e3D",
    "role": "BAL governance token"
  }
]

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
- https://defipunkd.com/address/1/0xA331D84eC860Bf466b4CdCcFb4aC09a1B43F3aE6 (Vault Authorizer)
- https://defipunkd.com/address/1/0x10A19e7eE7d7F8a52822f6817de8ea18204F2e4f (Balancer DAO Governance (Timelock))
- https://defipunkd.com/address/1/0x0EF2e7648A9b354822ed0EE9070A2E37332fd242 (Emergency SubDAO)
- https://defipunkd.com/address/1/0xA216683066513673FE033917bD5b11f871c65353 (Timelock)
- https://defipunkd.com/address/1/0xa731C23D7c95436Baaae9D52782f966E1ed07cc8 (Protocol fee controller)
- https://defipunkd.com/address/1/0xba100000625a3754423978a60c9317c58a424e3D (BAL governance token)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: OPEN-ACCESS

Evaluate who is allowed to use the protocol and whether any of that permission is granted off-chain.

Scope: this slice is about ADMISSION — who can enter, exit, or transact. Operator LIVENESS (what breaks if keepers/oracles go offline) is assessed in the dependencies slice and is out of scope for the grade here. You may note operator dependencies as context, but do not let "the protocol halts if operator X disappears" drive the access grade on its own; that belongs in dependencies. Source verification / contract verification on block explorers is assessed in the verifiability slice and is out of scope here — do NOT let "contract is unverified" drive the access grade.

Framing: the smart contracts are the access layer; frontends are UX. A permissionless contract is reachable by any client (SDK, third-party UI, aggregator, wallet integration). Frontend ToS, IP geo-blocking, and wallet screening are publisher policies on one specific client — they are reported as context but do NOT determine the grade. The grade hinges on (1) what the contract itself permits, and (2) whether the protocol is practically reachable without the official publisher's cooperation.

Meta-check before finalizing: if your verdict cites phrases from this prompt as evidence ("the protocol meets the 'credible alternatives' condition", "this fits the 'documented fallback' rule"), redo the verdict. The prompt describes the rubric; evidence must come from the protocol. A verdict should cite what the protocol does, not what the rubric says.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. Whitelist / allowlist modifiers in user-facing entry points. Grep for "onlyWhitelisted", "onlyRole", "allowlist", "isAccredited", "isKYCed". Note which functions are gated and who can add/remove from the list.
- A2. Off-chain operators in the admission path: keepers, sequencers, privileged relayers, oracle posters whose approval is required to admit a user action (not just to keep the protocol live). For each, identify whether the role is held by a single operator, a permissioned committee, or is permissionless. Enumerate per user-facing function class (deposit vs withdraw-request vs claim-finalized vs transfer) which ones require operator approval to be admitted, and which ones admit users unconditionally. A function whose placement is unconditional but whose downstream settlement depends on an operator is an admission-permissionless function — flag the liveness dependency as context and defer its grading weight to the dependencies slice.
- A3. Frontend restrictions on the official interface — record as context, not as a grade lever. Distinguish:
    - A3-passive: boilerplate ToS clauses (sanctions attestation, restricted-territory self-certification, VPN-circumvention prohibition, "comply with applicable law" eligibility, age of majority).
    - A3-active: runtime enforcement — IP-based geo-blocking, wallet-address screening against a sanctions oracle (Chainalysis, TRM, Elliptic), KYC wall, rendering-blocking jurisdiction banner.
  Record findings under the correct tier. Quote ToS text or banner text in evidence[].shows. These findings populate the headline and rationale but do NOT move the grade by themselves; the grade is set by A1, A2, and the A3b path check below.
- A3b. Independent access paths (the operative grade input). Enumerate paths that do not require the official publisher's cooperation:
    - Published SDK / library / CLI for direct contract interaction.
    - Third-party frontends operated by separate legal entities.
    - Wallet-integrated access (MetaMask Swaps, Safe apps, etc.).
    - DEX / lending / yield aggregators that route through the contracts.
  Record at least one concrete link per path that exists. The protocol does NOT have to self-document these — the test is existence, not UX cost. An A3b-i redistribution of the official UI bound by the same ToS does NOT count as an independent path.
- A4. Sanctions / compliance tooling at the contract level: does the protocol check addresses against OFAC lists or similar on-chain blocklists in the contract itself? (Frontend-only screening belongs in A3.)
- A5. Differentiate read access vs write access: many protocols are read-permissionless (anyone can view state) but write-gated (only certain addresses can deposit/borrow). Record both.
- A6. ToS / Legal links: locate them on the website and produce a VERBATIM quote of any jurisdictional, sanctions, or eligibility clause in evidence[].shows. If you cannot extract the clause text verbatim (SPA render failure, paywall, dead link, etc.), do NOT paraphrase or infer from general knowledge — record the ToS URL in unknowns[] with the reason extraction failed. Assertions about ToS content without a verbatim quote will be downweighted by reviewers.

Then write the steel-man section per Hard Rule 11.

Grade rules (admission-focused; liveness concerns belong in dependencies; source verification belongs in verifiability):
- green   = no contract-level whitelist/KYC on user entry/exit; no operator approval required to admit a user action; AND at least one independent A3b path exists (published SDK, third-party frontend, wallet integration, or aggregator routing). Frontend ToS posture and A3-active enforcement on the official UI do NOT block green when contracts are permissionless and an independent path exists — they are reported as context.
- orange  = contracts admit users unconditionally, BUT the protocol is operationally captured by the official publisher: no published SDK, no third-party frontend, no wallet integration, no aggregator routing. The contract is theoretically open but practically reachable only through the official UI. Also applies when admission requires approval from a permissioned committee that is governance-managed with a documented replacement procedure.
- red     = contract-level whitelist / KYC on user entry/exit, OR admission of a core user action requires approval from a single privileged operator or a small committee with no documented replacement procedure, OR enforces an on-chain blocklist updatable by a single party.
- unknown = checklist incomplete after checking the sources above.

Default-grade guidance: when contracts are fully permissionless AND any A3b independent path exists, the default grade is green regardless of frontend ToS or A3-active enforcement on the official UI. Frontend geo-blocking, sanctions-oracle wallet screening, and ToS sanctions clauses are publisher policies on one client and are reported in findings/headline as context, not as grade levers. To grade orange on operational-capture grounds, the auditor must affirmatively show that ALL independent paths are absent or also gated.

Guideline on committees: where admission depends on a multi-operator committee, the relevant axes are (a) set size, (b) whether replacement/rotation is governed on-chain, (c) whether the replacement procedure is publicly documented. A large set with on-chain governance replacement should not be graded as a single-party operator even if rotation is not instantaneous. A small set with informal replacement should be treated as a single-party operator.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "open-access",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Sources gpt-5.5 no url gemini-3-flash-preview no url View raw submissions ↗

Stage

Preview of the Phase-3 maturity framework. DeFiPunk'd will adopt DeFiScan v2's stages verbatim; the section is rendered below in its intended shape so the structure is visible today.

Balancer V3 has not yet been assessed under the DeFiScan v2 stage framework.

The walkaway test is the central criterion. Once stages land, protocols reach Stage 1 only if users can exit in the presence of malicious operators even when the emergency council disappears.

Scope of assessment

Stages are assessed per-protocol against DeFiScan v2's criteria: governance structure, upgradeability path, timelock durations, emergency-council scope, and the walkaway test. The analysis depends on onchain discovery (roles, owners, timelocks) and deeper review of deployed contracts — neither of which DeFiPunk'd automates at Phase 0.

Stage 0 requirements pending

Governance is largely off-chain, contracts are upgradeable with short or no timelock, and the protocol depends on a multisig or team with full discretion. At Phase 0 DeFiPunk'd does not automatically evaluate these; the assessment lands with crawler-based onchain discovery.

Stage 1 requirements pending

Users can exit or opt out on their own terms even if the team disappears. Upgrades run through a meaningful timelock with an emergency security council clearly scoped. The walkaway test is the headline criterion.

Stage 2 requirements pending

Protocol is fully permissionless and immutable, or upgrades require a supermajority of token holders with a long timelock and no emergency override. This is the terminal stage of the DeFiScan v2 framework.

Learn more about DeFiScan v2 stages →

Stages are an opinionated assessment of maturity, not a rating of security or safety. A protocol can sit at Stage 2 and still carry substantial technical or economic risk; the framework exists to incentivize decentralization, not to rank protocols.

TVL $30.5M

Type Dexs

Chains Ethereum, Polygon, Arbitrum, Base, xDai +4

View on DeFiLlama ↗

Control Exit Autonomy Open Access Verifiable

Verifiability tentative Open source + 3 audits
Control unknown Not yet assessed
Ability to exit unknown Not yet assessed
Autonomy unknown No Phase-0 autonomy signal
Open Access unknown Not yet assessed

Control criteria

Upgradeability — Bug bounty immunefi.com Governance forum — Docs —

About

Balancer V2 is a dexs protocol on Ethereum.

Risk analysis

One card per dimension, sorted by severity. Only Verifiability and Autonomy carry automated signals in Phase 0. See methodology for scope.

Audit a dimension yourself · DEFI@home Contribute an LLM-run assessment — any model, any dimension. Three agreeing runs merge automatically into the public record.

Address discovery 0 addresses on file · 0 runs ⚑ Run first Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v2
- protocol.name:              Balancer V2
- protocol.chains:            Ethereum, Polygon, Arbitrum, Base, xDai, Avalanche, Polygon zkEVM, Mode, Fraxtal
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/balancer/balancer-v2-monorepo/tree/master/audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: DISCOVERY

You are a cataloguer, not a judge. Your job is to surface every contract address that could plausibly belong to this protocol's control or fund-holding surface, each backed by a citation. `grade` is ALWAYS `"unknown"` for discovery submissions — there is no green/orange/red rubric here. The five evaluation slices that run after you (control, ability-to-exit, autonomy, open-access, verifiability) consume your output via the addressBook ratchet — every address you record becomes a pre-built surfacer URL on the next run; every address you miss costs them a tool call.

Width beats depth. A `role: "other"` entry with one cited URL beats omitting it. Downstream slices will discard out-of-scope entries; they cannot rediscover what you fail to enumerate without paying the same cost again.

(Step 0 capability probe lives in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every D-code below must appear in evidence[] OR unknowns[])

- **D1. Block-explorer name-tag search per chain.** For each chain in `protocol.chains`, search the canonical block explorer for the protocol's name tag — `https://etherscan.io/searchHandler?term=<query>` and the per-chain explorers (basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network). When direct fetch is blocked, use `site:<explorer> <protocol_name>` via search grounding. Record every address that surfaces with the protocol's tag, plus neighbouring "Token Contract" / "Multisig" / "Timelock" labels.

- **D2. Official deployments doc.** From `protocol.website` and the docs site, locate the canonical "Deployed contracts" / "Addresses" / "Contracts" / "Deployments" page (often at `/docs/deployments`, `/docs/addresses`, `/dashboard/contracts`). Cite the URL, record every address listed with its named role, and set `protocol_metadata.deployed_contracts_doc` to this URL.

- **D3. Audit PDFs.** From `protocol.audit_links` (and any audits surfaced by D2), open each. Most reports include a "Scope" / "Contracts in scope" address table in the first 5 pages. Extract every in-scope address with its labelled role. If the audit predates the current deployment, record the addresses anyway with role suffixed `(audit-era)` so downstream slices know to re-verify.

- **D4. GitHub deployment artifacts.** From `protocol.github`, walk the repo at a pinned commit SHA looking for: Foundry `broadcast/<script>/<chainId>/run-latest.json` (`transactions[].contractAddress` per chainId); hardhat-deploy `deployments/<network>/<Contract>.json` (`address` field); manual indexes (`deployments.json`, `addresses.json`, `contracts.json`, `networks.json`); markdown indexes (`docs/deployments.md`, `README.md` tables). Cite the file URL with the commit SHA; pin SHAs (`?ref=<sha>`) so the citation is content-addressed.

- **D5. Multi-chain enumeration.** If `protocol.chains.length > 1`, repeat D1–D4 per chain. Cross-chain deployments of the same logical contract get SEPARATE `admin_addresses[]` entries — one per chain. The chain field is part of the identity; do not collapse. If a chain has zero results, record `"D5: chain <name>: zero addresses surfaced from <sources tried>"` in unknowns[].

- **D6. Factory-discovered children.** For factory addresses surfaced in D1–D4, fetch the enumeration view via the read API (`/api/contract/read?...&method=allPools` / `getPool` / `getMarket` / `getVault`) and record each child with role like `"pool (from factory <0xFactory>)"`. **Cap at 50 children per factory.** Protocols with thousands of pools (Uniswap, Sushi) need dedicated ingestion — record the factory + the cap notice in unknowns[].

- **D7. Role taxonomy.** Every `admin_addresses[]` entry's `role` uses this controlled vocabulary (free-text suffixes OK for disambiguation, e.g. `"multisig (treasury)"`, but the leading token must match):

    `owner | admin | proxy_admin | governor | timelock | guardian | multisig | treasury | oracle | factory | router | token | pool | vault | other`

  Tentative classifications are encouraged. `actor_class` ∈ `eoa | multisig | timelock | governance | unknown` — use `unknown` when you found the address but didn't read its bytecode.

- **D8. Ratchet output integrity.** Every address in `admin_addresses[]` must trace to ≥1 fetched URL in evidence[]. Snippet-only sightings go in unknowns[] with a `D8` code, NOT in admin_addresses[].

### Discovery rationale framing

- `rationale.findings`: one entry per D-code, terse, factual. Per-address detail belongs in evidence[] and admin_addresses[], not here. Example: `"D1: 8 addresses surfaced from etherscan.io name-tag search for 'Aave V3'"`.
- `rationale.steelman`: ALWAYS null.
- `rationale.verdict`: one short line summarizing what corpora were walked and how many addresses were catalogued.
- `headline`: factual and quantitative — `"24 contracts catalogued across Ethereum, Arbitrum, and Base; 6 governance/admin and 18 protocol contracts."`.
- `short_headline`: under 60 chars — `"24 contracts across 3 chains"`.

### What discovery is NOT

- Not a verdict slice. `grade` must be `"unknown"`.
- Not exhaustive enumeration of leaf assets — record the factory + cap and move on (see D6).
- Not classification of trust assumptions — whether a multisig threshold is safe / timelock delay is sufficient / proxy admin is an EOA is the control slice's job.
- Not address-book reconciliation: when addressBook is non-empty, EXTEND it (find addresses prior runs missed) rather than re-cite the same addresses; re-cite only when you have new evidence for a refined role.

### protocol_metadata side-effects

While walking the corpora, populate every `protocol_metadata` field you can support with citations: `github`, `docs_url`, `audits` (one per D3 audit walked), `governance_forum`, `bug_bounty_url`, `security_contact`, `deployed_contracts_doc` (URL from D2), `upgradeability` (best-effort), and `about` (2–4 sentences sourced from docs/website, not memory). Discovery is the natural home for these — evaluation slices should not have to rediscover them.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "discovery",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Verifiability Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v2
- protocol.name:              Balancer V2
- protocol.chains:            Ethereum, Polygon, Arbitrum, Base, xDai, Avalanche, Polygon zkEVM, Mode, Fraxtal
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/balancer/balancer-v2-monorepo/tree/master/audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: VERIFIABILITY

Evaluate whether an outsider can independently confirm what the deployed code does.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- V1. For each address you assess: is the bytecode verified on the chain's block explorer? Record the "Contract Source Code Verified" indicator. https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x... returns this as a top-level "verified" boolean plus "abiSource" ("etherscan" / "sourcify") and an inline ABI — useful when the explorer page is rate-limited. If the contract is a proxy, verify BOTH the proxy contract AND the current implementation contract. The same /api/contract/abi response auto-resolves proxies and includes a "proxy.implementation" address when present, so one call covers V1 + V6 in one shot. An explorer "Similar Match" on a well-known proxy pattern (Aragon AppProxyUpgradeable, ERC1967Proxy, OssifiableProxy, OZ TransparentUpgradeableProxy) is expected for that pattern and does NOT count as a verification gap on its own — what matters is that the implementation is independently verified.
- V2. Source-to-repo correspondence: for each verified contract, attempt to find a matching commit in the linked GitHub repos. Record evidence[].commit on a match. Independent compile/bytecode-match is NOT required for green — a recognized public repo whose structure and file contents correspond to the explorer-visible source is sufficient. If you did not pin a commit SHA or run a bytecode diff, record that plainly in unknowns[] and proceed; it is a scope limit, not a downgrade signal.
- V3. Audit coverage: for each URL in protocol.audit_links, open it and record: auditor name, audit date, the specific contracts / commit in scope. Flag audits that predate the current deployment by >6 months without a follow-up review.
- V4. Auditor recognition: the following firms are broadly recognized in Solidity: Trail of Bits, Zellic, Spearbit, OpenZeppelin, ConsenSys Diligence, Certora (formal verification), Quantstamp, Halborn, Peckshield, Sigma Prime, ChainSecurity, Ackee Blockchain, MixBytes, Statemind. Unknown firms are orange-at-best for any green-grade claim. Name the firm explicitly in evidence[].
- V5. Post-audit drift: compare the most recent in-scope audit(s) against the currently-deployed source, weighted by what each contract does and by what the changes actually contain. SCOPE — drift only downgrades the grade when ALL of the following hold: (i) the drifting contract is fund-custody / settlement / accounting-critical (NOT a peripheral router, lens, quoter, or pure-view contract that holds no balances); (ii) the changes are material — new functions, modified access control, modified accounting, modified fund flow — and not refactors / struct relocations / import reorgs / build or CI fixes / formatting; (iii) no later audit, fix-audit, or differential audit from a recognized firm covers the changed files (audits often pin a pre-fix commit while a follow-up reviews the delta — match by file scope, not by commit-hash equality). When you cite drift as a downgrade reason, name the specific behavior change (function added, role granted, accounting formula altered) — "N commits ahead" or "+X/-Y LOC" alone is not evidence of material drift. If you have not sampled the diff content (e.g. via the GitHub compare view or the top commits in the window), record drift as an unknowns[] entry rather than auto-downgrading; commit-count and LOC are starting signals, not findings.
- V6. Implementation vs proxy: a verified proxy with an unverified implementation is effectively unverified. State whether the implementation is verified separately.

EVIDENCE DISCIPLINE (read before writing findings[]):
- Do not assert a specific deploy-commit SHA, bytecode equivalence, or "identical to audited commit" unless you actually fetched the artifact that shows it (e.g., a deployed-addresses JSON you opened, an explorer page you read). Inferred or plausible matches belong in unknowns[], never in findings[] or evidence[].
- Evidence[] entries must correspond to pages/files you actually retrieved this run. A URL you did not open is not evidence.

Then write the steel-man section per Hard Rule 11.

Grade rules:
- green   = deployed bytecode verified on the explorer (proxy AND implementation if proxied; "Similar Match" on a standard proxy pattern is fine per V1), a public source repo exists whose contents correspond to the explorer-visible source, AND ≥1 audit from a recognized firm covering the currently-deployed contracts (≤6 months of drift OR drift was re-audited). A missing local compile-match is not a downgrade — record it in unknowns[] and still grade green if the other conditions hold.
- orange  = verified but with visible drift from the public repo, OR audit scope is stale relative to deployment, OR only minor / unknown-firm audits exist, OR only some of the main contracts are verified, OR proxy verified but implementation only partially verified.
- red     = unverified bytecode (or verified proxy with unverified implementation), OR no audit in protocol.audit_links, OR no public repo.
- unknown = reserved for when the protocol's verifiability posture genuinely cannot be assessed (e.g., explorer and repo both inaccessible for this protocol). Do NOT use unknown merely because you, the analyst, could not run a particular check such as a bytecode diff — that goes in unknowns[] while the grade is still assigned from the evidence you do have.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "verifiability",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Control Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v2
- protocol.name:              Balancer V2
- protocol.chains:            Ethereum, Polygon, Arbitrum, Base, xDai, Avalanche, Polygon zkEVM, Mode, Fraxtal
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/balancer/balancer-v2-monorepo/tree/master/audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: CONTROL

Evaluate who can change the protocol's rules, how fast, and how broadly.

(Step 0 capability probe and the off-chain-only fallback live in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[])

- **C1.** For each address you assess: who is the contract owner / admin / pendingAdmin / governor — read these via the block explorer's "Read Contract" tab OR `https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=owner` (BARE method names: `&method=owner`, `&method=admin`, `&method=pendingOwner`, `&method=governor`). For Safes use `/api/safe/owners?chainId=<id>&address=0x...`. When a protocol has multiple major versions deployed (v2/v3/v4), perform C1 reads on the NEWEST deployment separately — newer deployments often have weaker control surfaces than the legacy core.
- **C2.** Upgrade mechanism: transparent proxy / UUPS / Beacon / Diamond / immutable. Identify the proxy admin address. Check upgradeability of GOVERNANCE contracts too — a Governor / Aragon Voting / OZ Governor is often itself a proxy whose admin is the Timelock. Asymmetry: when fund-holding cores are immutable AND governance has no admin path that reaches them, an upgradable Governor/Timelock is T3-only and must NOT drag the verdict below green on that basis alone (see grade rules and the "immutable cores" caveat). Only call upgradability "mixed" if you can name a concrete function on the upgradable surface that reaches T1 or T2 on user funds.
- **C3.** EXECUTION PATH (enumerate every stage, in order, with delays in seconds). The operative path is usually a chain (voting → scheduler → timelock → executor; or governor → queue → execute; or Aragon Voting → DualGovernance → EmergencyProtectedTimelock → AdminExecutor). For each stage, record (a) the contract address, (b) the delay constant name + value in seconds, (c) the URL you read it from (block-explorer Read Contract OR `/api/contract/read?...&method=MIN_DELAY&block=<n>`). Do NOT stop at the first timelock-shaped contract — if its admin is itself called by another contract, keep walking. The grading delay is the SUM OF DELAYS ON THE UNCONTESTED FAST PATH (shortest time a proposal with no opposition can go from submission to executable). Dynamic / contested extensions (veto signaling, rage quit, escrow delay) are modifiers, not the basis — note them separately.
- **C4.** Enumerate EVERY multisig with reachable control — main proxy admin, emergency activation, emergency execution, reseal / pause, gate-seal committees, tiebreaker, per-module admins. For each Safe, fetch threshold + owners + version via `/api/safe/owners?chainId=<id>&address=0x...` (response includes raw eth_call data, so the URL is citable evidence). Enumerate ops/council/incentives multisigs even when off the upgrade path — record their scope so a reader can see they are NOT on the upgrade path. For each: (a) address, (b) threshold / total signers, (c) signer identities classified as insider (team, paid auditors under ongoing engagement, mandated service providers) vs non-insider (independent community members, unaffiliated researchers), (d) the specific power held (upgrade, pause, parameter, etc.).
- **C5.** On-chain governance: Governor / GovernorBravo / OZ Governor / Aragon Voting with token-weighted voting? Record proposal threshold, voting period, quorum, and the timelock delay between queue and execute. Every numeric constant must come from a Read Contract call you can link to, or be in unknowns[] with the C-code. If votingDelay / votingPeriod are denominated in BLOCKS, convert to seconds at the chain's CURRENT block time (Ethereum mainnet ≈ 12s post-Merge, not the 15s in older Compound/Bravo deployments) — cite both block count and converted seconds.
- **C6.** EMERGENCY POWERS: separate emergency-pause / guardian role with a different time cap or different actor than the main upgrade authority? Record it explicitly.
- **C7.** POWER TIER (blast radius). For each privileged path in C3–C6, classify the WORST thing that path can do, choosing the highest applicable tier. Cite the specific function name and any on-chain bound — tier claims without a named function are unsupported.
    - **T1 — FUND-CRITICAL**: replace implementation of contracts holding user funds; change AMM math / accounting / collateral logic; mint unbacked debt or shares; pause withdrawals; drain user-fund treasury; change oracle to attacker-controlled source; replace upgrade admin with EOA.
    - **T2 — ECONOMICALLY MATERIAL**: change fee parameters within bounded ranges; redirect protocol fees; add/remove markets / collateral types; bounded inflation or token mint within hard-capped schedule; spend protocol-owned (non-user) treasury.
    - **T3 — GOVERNANCE-INTERNAL**: change voting rules, quorum, voting period, proposal threshold; upgrade the Governor itself; rotate Timelock admin; mint governance tokens within a capped annual schedule.
    - **T4 — OPERATIONAL**: incentives distribution, grants, ENS / frontend canonicalization, deployment coordination, periphery router deprecation.
  The grade is set by the HIGHEST tier reachable on the uncontested fast path, not the median. State the tier and the binding function in the verdict.

### Read Contract discipline (applies to C3, C4, C5)

Every numeric constant cited (timelock delays, voting periods, multisig thresholds, quorum percentages) must come from EITHER (a) a block-explorer Read Contract URL, OR (b) a DeFiPunkd `/api/contract/read` or `/api/safe/owners` URL (preferred with `&block=<n>` for content-addressed evidence), OR (c) an unknowns[] entry with the C-code. Docs / blog posts are corroboration only — they cannot be the sole citation for a value that is also readable on-chain.

### Off-chain-only substitute hierarchy (when grading_basis="off-chain-only" — see preamble Rule 16)

When on-chain reads were genuinely unreachable this run, eligible off-chain substitutes in priority order:
1. Linked audit PDFs (admin roles, multisig members, timelock delays usually enumerated).
2. Governance forum posts that quote constants from a successful on-chain proposal (cite post URL + linked execution-tx URL).
3. Official protocol docs pages with named addresses and roles (must be on a domain owned by the protocol).
4. GitHub README / SECURITY.md / governance/*.md at a pinned commit SHA.

Forbidden substitutes: third-party blog posts, X / Twitter threads, search-result snippets, model memory. Required degradation: any C-code citing a numeric constant from docs/forum/audit prose ONLY must also carry an `unknowns[]` entry with `-offchain` suffix noting "value not re-read on-chain in this run; corroboration only".

### Grade rules (apply the timelock bar conditional on the highest C7 tier reachable on the fast path)

Security Council standard (used below): a multisig qualifies as "Security Council" only if ALL of: ≥7 signers, ≥51% threshold, ≥50% non-insider signers, every signer publicly announced. Failing any criterion = NOT a Security Council, regardless of signer reputation.

- **green**: highest reachable tier is T3 or T4 regardless of timelock; OR T2 reachable with uncontested-fast-path delay ≥7 days; OR T1 reachable only via immutable contracts (T1 is unreachable); OR T1 reachable with uncontested-fast-path delay ≥7 days combined with a Security Council multisig; OR T1 reachable with uncontested-fast-path delay ≥7 days through active on-chain governance with broad token distribution.
- **orange**: T2 reachable with uncontested-fast-path delay >0 but <7 days; OR T1 reachable with uncontested-fast-path delay >0 but <7 days; OR a multisig failing one or more Security Council criteria sits on a T1/T2 path; OR unclear upgrade authority on a T1/T2 path; OR governance with very short timelock or low quorum on a T1/T2 path.
- **red**: T1 reachable with no timelock by a single EOA or 2-of-3 multisig; OR a T1 upgrade admin that is not a smart contract you can audit.
- **unknown**: completed the checklist but still cannot determine the upgrade authority OR cannot classify the highest tier reachable on the main contracts.

Tiering caveats:
- "Bounded" must be enforced ON-CHAIN to count as T2. A function that sets fees with no upper-bound check is T1 — cite the bound check.
- Recurring T2 economic extraction (e.g. fee redirect with no rate limit) approaches T1 over time. A single proposal that can permanently redirect all future revenue is T1.
- T3 assumes the governance contract cannot itself authorize a T1/T2 action without going through the same timelock. If governance can self-upgrade to bypass the timelock, T3 collapses into T1.
- Do not downgrade tier by hand-waving ("realistically governance would never…"). Tier on what the contract permits, not what feels likely.

Notes:
- **Dynamic / dual-governance timelocks** (Lido, Compound escrow veto): the rubric grades on the uncontested path because that is the path most upgrades take. A dynamic extension that fires only under stake-weighted opposition is a real protection — name it in the green steel-man, but it does not lift an orange fast path into green; state the tension in the verdict.
- **Immutable cores with upgradable governance** (Uniswap-style): if fund-holding contracts are immutable and have no admin-reachable function moving / freezing / re-routing user funds, the highest reachable tier on the upgrade path is T3 — green regardless of timelock. Don't grade this orange just because the Governor is a proxy — that's a C2 fact, not a downgrade. Downgrade only applies if you can cite a concrete function on the upgradable surface that reaches T1 or T2 (privileged hook, upgradable factory controlling fund-routing, fee-switch redirecting protocol revenue without bound).
- **The 7-day bar** reflects the exit-window standard — users need notice after a queued upgrade to withdraw if they disagree. The ability-to-exit slice grades the exit side; this slice grades the delay side; both must hold for users to actually benefit from the delay.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "control",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Ability to exit Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v2
- protocol.name:              Balancer V2
- protocol.chains:            Ethereum, Polygon, Arbitrum, Base, xDai, Avalanche, Polygon zkEVM, Mode, Fraxtal
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/balancer/balancer-v2-monorepo/tree/master/audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: ABILITY-TO-EXIT

Evaluate whether users can withdraw their funds on their own terms, even under adversarial admin conditions.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- E1. Enumerate every user-facing exit function in the main contracts: withdraw, redeem, burn, requestWithdrawal, claim, exit, etc. List them by name. Do NOT treat the contract as a monolith.
- E2. For EACH exit function in E1: identify its access modifiers and any pause guards (e.g. _checkResumed, whenNotPaused, onlyRole). Functions that gate REQUEST PLACEMENT often differ from functions that CLAIM FINALIZED FUNDS — check both separately.
- E3. For each pause guard: identify the role holder (which address holds PAUSE_ROLE / GUARDIAN / etc.) and the maximum pause duration. Specifically check whether PAUSE_INFINITELY (or equivalent uncapped pause) is callable, and which actor can call it (single multisig vs governance vote). For role-holder reads use https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=hasRole&args=0x...,0x... or &method=getRoleAdmin&args=0x.... For "is currently paused" checks use &method=paused or &method=isPaused&args=<resume-code>. Use the BARE method name (no parens). Cite the URL with &block=<n> in evidence[].
- E4. EMERGENCY vs GOVERNANCE pause distinction: many protocols have a fast-acting emergency pause capped at N days and a slow governance pause that can be indefinite. Record both paths separately if present, with their time caps and actor classes.
- E5. Queued redemption: documented maximum queue duration, daily withdrawal caps, whether the queue itself is pausable.
- E6. Forced-exit / escape-hatch / permissionless emergency-exit mechanism for adversarial-admin scenarios.
- E7. Frontend dependency: confirm exit functions are directly callable on-chain (e.g. via Etherscan write tab or a generic wallet) without the project's frontend.

Then write the steel-man section per Hard Rule 11. Common red-vs-orange tension on this slice: indefinite pause exists (suggests red) BUT the realistic emergency path is time-capped AND claims of already-finalized exits are not pause-gated (suggests orange). Resolve this by stating who can do what for how long, not by stopping at the worst-case sentence.

Grade rules:
- green   = permissionless exit; pause is either absent, narrowly scoped to clearly-described emergencies with auto-expiry, or capped at ≤7 days; no frontend dependency for exit; claims of already-finalized exits are not pause-gated under any path.
- orange  = pausable with broad scope OR indefinite pause is reachable only through governance vote (not unilateral admin action), OR queued redemption with documented max > 7 days, OR claims-of-finalized are exempt but new-request placement can be paused indefinitely by governance.
- red     = exit requires admin signature, OR ANY actor (including governance) can pause CLAIMS of finalized exits indefinitely, OR there is no on-chain exit function at all (purely custodial), OR pause is held by a single EOA / 2-of-3 multisig with no time cap.
- unknown = checklist incomplete after checking the sources above.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "ability-to-exit",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Autonomy Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v2
- protocol.name:              Balancer V2
- protocol.chains:            Ethereum, Polygon, Arbitrum, Base, xDai, Avalanche, Polygon zkEVM, Mode, Fraxtal
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/balancer/balancer-v2-monorepo/tree/master/audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: AUTONOMY

Evaluate this protocol's autonomy: can a failure of anything outside its own contracts cause theft or loss of user principal, loss of unclaimed yield, or materially change the protocol's expected performance? "Autonomy" is not "has zero external touchpoints" — it is "external failures are either survivable or recoverable without user loss". This slice adapts DefiScan v1's Autonomy dimension; dependencies are one of several criteria, not the whole frame.

GUARDRAILS (read before grading):
- Category alone (Liquid Staking, Bridge, RWA Lending, Restaking, …) does NOT force a grade. A category is a hint about where to look; the grade must come from the concrete A1–A9 findings below.
- Base-chain consensus (Ethereum PoS, the chain's validator set, the canonical Deposit Contract at 0x00000000219ab540356cBB839Cbe05303d7705Fa) is the SUBSTRATE, not a dependency, for any protocol deployed on that chain. Do not list "depends on Ethereum" as a finding.
- Oracles or other integrations used by DOWNSTREAM protocols that happen to read this protocol's token (e.g. Chainlink stETH/USD consumed by Aave) are NOT this protocol's dependencies. Count only what THIS protocol's contracts call or trust on-chain. If a feed or contract appears in the protocol's docs only as reference material for third-party integrators, EXCLUDE IT ENTIRELY — do not log it as a finding even with a "peripheral" or "referenced only" caveat.
- "Upgradeable admin can change things" belongs to the CONTROL slice; only count it here when the upgrade surface lets governance silently swap an external dependency (see A9).
- Underlying-asset risk in opt-in, isolated markets is NOT autonomy-red on its own. When a protocol wraps third-party yield-bearing assets (LSTs, LRTs, ERC-4626 vaults, lending receipts, restaked tokens) into per-market silos that users explicitly choose, a failure of one underlying does NOT propagate to other markets and is risk the user opted into per-market. Record it under A4 with depth and propagation scope, but do not let it alone drive a red verdict — red requires an external dependency that cross-cuts the protocol or that the user did not opt into at deposit time. A failure mode that is "if the LRT you deposited is hacked, your principal in that LRT-backed market is impaired" is the underlying's autonomy story, not this protocol's; grade it on whether THIS protocol introduces additional dependencies on top.
- Sub-module enumeration is mandatory before grading. If the protocol ships distinct product lines or modules (e.g., a v2 core AMM plus a newer perps/funding-rate module, a lending pool plus a separate vault layer, an L1 core plus a cross-chain extension), enumerate each in findings and grade against the WORST module weighted by its share of TVS or its blast radius. A green core does not rescue an orange/red sub-module; conversely, a small red sub-module with capped TVS may bound the overall grade to orange. Name each module by its on-chain factory or router address. If you do not know whether a module exists, that is an unknowns[] entry, not silence. EXPLICIT TVS WEIGHTING: in the verdict, state each module's approximate share of total protocol TVS (use "~X%" if exact figures unavailable; check DeFiLlama or block-explorer balances on the module's main contract) and how that share informs the weighted grade. Format: "Module A holds ~X% of TVS (grade: <g>); Module B holds ~Y% (grade: <g>); weighted overall = <grade> because <reason>." If a red sub-module holds <5% of total TVS and is capped, the overall grade may be orange; if it holds >25%, the overall grade is red. Do not let qualitative reasoning substitute for the percentages — write the numbers.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. External contract calls. Enumerate every external contract the core contracts call or read from (oracles, price feeds, AMM pools, lending pools, staking/deposit contracts, yield wrappers). For each, identify the address, the provider, and what user-facing function of this protocol would break or mis-price if that external contract paused, mis-reported, or behaved adversarially. Grep for "oracle", "aggregator", "getPrice", "latestAnswer", "chainlink", "pyth", "redstone" as a starting point. To verify an oracle is live and what it currently reports, hit https://defipunkd.com/api/contract/read?chainId=<id>&address=<oracle>&method=latestAnswer (or &method=latestRoundData, &method=getPrice, &method=description; use the BARE method name without parens) and cite the URL — the response includes blockNumber/blockHash and rawReturnData, which is stronger evidence than a docs page about which feed "should" be used.
- A2. Off-chain actor committees reporting INTO the protocol. Oracle committees, guardian multisigs, DAO-selected validator sets acting as protocol reporters, exit-bus signers, fraud-proof challengers. For each, record committee size, quorum, who picks the members, and what mis-reporting could do (mint, burn, finalize withdrawal, freeze). Distinguish this from governance admins (control slice). NOTE ON STAKING PROTOCOLS: validator slashing and node-operator misbehavior are properties of the base-chain substrate, NOT external dependencies, when the operator set is diversified enough that a coordinated failure caps at <5% principal loss. Count validator/operator risk under A2 ONLY if the operator set is small, non-diversified, or lacks bonding / slashing-insurance / diversification mitigations. A curated set of 30+ independent operators with documented diversification falls under the mitigated path; a 3-operator LST does not. Do not cite the protocol's own risk disclosure as evidence that operator failure = principal loss unless you also check the diversification and bond mitigations.
- A3. Bridge / cross-chain messaging dependencies. Only count bridges that carry material TVL or are required for a core user flow. For each, name the bridge operator (canonical L1↔L2, LayerZero, Wormhole, Axelar, custom multisig), the trust model (canonical, optimistic, light-client, guardian set), and what fraction of TVL or users ride on it. "wstETH exists on 15 chains" is not a finding unless material TVL sits there. Before listing any non-primary chain deployment as a dependency, verify it is still operational as of analysis_date — retired or sunset deployments (e.g., Lido-on-Terra, Lido-on-Solana) belong in unknowns[] or should be omitted, never cited as a current dependency.
- A4. Nested collateral / restaking chains. For restaking / LRT / receipt-of-receipt designs: record the depth of the collateral chain, every actor with slashing or freezing power at each level, and whether a failure N levels deep propagates to user principal here.
- A5. Fork lineage (silent check). If DeFiLlama's forkedFrom is non-empty, record it as one finding and move on. If empty, do NOT add a placeholder finding; it adds noise.
- A6. Fallback mechanisms and circuit breakers. What catches an external failure? Sanity-check contracts on oracle reports, rebase bounds, pause paths triggered by bad prices, second-opinion oracles, max-per-block throughput caps, withdrawal queues that absorb bad reports. Record which A1–A4 risks are mitigated by which fallback, and which are unmitigated. For EACH fallback you cite, state its activation status explicitly: (i) LIVE and enforcing on-chain today, (ii) DEPLOYED but not yet wired / activated (e.g., interface exists but the address is zero or the role is unassigned), or (iii) DOCUMENTED / PROPOSED only (forum post, LIP draft, audit pending). Only (i) counts as mitigation for the grade. A fallback in state (ii) or (iii) should be noted but must not reduce the risk in your steel-man or verdict. If you cannot determine activation status, add an unknowns[] entry rather than assume it is live.
- A7. Sequencer / L1-liveness dependency BEYOND the base-chain substrate. SCOPE — sequencer risk only counts here when the protocol IS its own L2/L3 appchain or app-rollup, where the sequencer is part of the protocol's own stack and a freeze is a protocol-level outage. A protocol permissionlessly deployed on a third-party L2 inherits that L2's sequencer as substrate, not an A7 dependency. Record the sequencer/DA trust model when A7 applies.
- A8. Keeper / relayer / off-chain bot liveness. Protocols that need permissionless-but-necessary off-chain actors (liquidation bots, auto-compounders, deposit relayers, intent solvers). Record whether the role is permissionless, what degrades if nobody runs it (yield paused, bad debt accumulates, positions go stale), and whether the failure mode is graceful or catastrophic.
- A9. Governance-mutable dependency surface. Can an admin or DAO action silently INTRODUCE a NEW EXTERNAL dependency — swap the oracle address to a different provider, register a new staking module that calls an untrusted contract, add a new bridge, route SY through a new external vault — without an exit window for users? Check the upgrade / router / module-registry contracts. Answer: which EXTERNAL dependencies are governance-mutable, who holds that power, and whether there is a timelock or exit window. SCOPE LIMIT — read carefully: A9 is about the *external dependency surface*, not the upgrade surface in general. "The proxyAdmin / EOA can upgrade the router implementation to arbitrary bytecode" is a CONTROL-slice finding (admin can rug), NOT an autonomy-A9 finding. A9 fires only when the upgrade specifically swaps out or adds a contract that THIS protocol calls or trusts (e.g., changing the oracle address from Chainlink to a malicious feed, registering a new SY adapter that points to a third-party vault, redirecting a bridge endpoint). If the only finding is "admin can change implementation," do not log it under A9 and do not let it drive the autonomy grade — note it under control instead. The autonomy-relevant version of the same upgrade key is "admin can swap [specific external address X] without timelock"; that requires identifying the specific external dependency that becomes mutable.

STEEL-MAN (per Hard Rule 13): write one-sentence strongest arguments for red, orange, and green using the A1–A9 findings.

IMPACTED TVS ESTIMATE: the headline MUST include a rough impacted-TVS figure — the fraction of protocol TVS that could be lost or frozen if the worst-unmitigated dependency you identified failed. Use "~X%" if exact numbers are unavailable, "<1%" for de minimis, "unclear" only if A1–A9 left the question genuinely open (in which case grade=unknown is usually correct). Do NOT substitute qualitative phrases like "significant" — give a number or bracket.

GRADE ANCHORS (mapped to DefiScan v1 stages):
- green = Stage 2 equivalent. Failure of any external dependency cannot cause loss of user principal or unclaimed yield. Either there are no material external dependencies, or every critical one has a documented fallback (A6) that keeps users whole. Governance cannot silently introduce new dependencies without an exit window (A9). Impacted TVS under any single-dependency failure: effectively 0.
- orange = Stage 1 equivalent. Failure of some external dependency can cause loss of unclaimed yield, or can materially change expected performance (pause withdrawals, freeze positions, degrade price quality), but cannot cause loss of principal. Committee-based oracles with sanity checks, canonical-only bridges, fallback paths that exist but are incomplete, or governance-mutable dependencies protected by a ≥7-day timelock. Impacted TVS is bounded and recoverable.
- red = Stage 0 equivalent. Failure of an external dependency CAN cause theft or loss of principal. Examples: single-provider oracle with no sanity check or fallback, material TVL on a non-canonical bridge with a guardian multisig, governance can hot-swap oracles or add staking modules with no timelock or exit window, unmitigated keeper-liveness dependency where positions become insolvent if bots stop. Impacted TVS is material.
- unknown = checklist incomplete after inspecting source + verified contracts. Prefer unknown over guessing when A1/A6 could not be reconstructed. RESERVE unknown for cases where the CORE ARCHITECTURE itself is unverifiable — not for cases where you merely cannot enumerate every per-market dependency in a multi-market protocol. If the core router/factory/oracle architecture is verifiable on-chain and you can determine whether the core requires external dependencies, grade the architecture even when an exhaustive per-market external-dependency census is infeasible. Acknowledge the per-market gap in unknowns[] but still issue a grade. Refusing to grade a multi-market protocol because you cannot list every SY/vault/market is over-use of unknown; grade the architecture and say so.

PROMPT-META CHECK (per Hard Rule 17): before finalizing, verify the verdict cites concrete contract addresses, docs, or code — not the rubric itself. If your verdict says "the protocol belongs to a category the rubric marks red", rewrite it with the A1–A9 finding that actually justifies the grade, or drop to grade=unknown.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "autonomy",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Open Access Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v2
- protocol.name:              Balancer V2
- protocol.chains:            Ethereum, Polygon, Arbitrum, Base, xDai, Avalanche, Polygon zkEVM, Mode, Fraxtal
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/balancer/balancer-v2-monorepo/tree/master/audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: OPEN-ACCESS

Evaluate who is allowed to use the protocol and whether any of that permission is granted off-chain.

Scope: this slice is about ADMISSION — who can enter, exit, or transact. Operator LIVENESS (what breaks if keepers/oracles go offline) is assessed in the dependencies slice and is out of scope for the grade here. You may note operator dependencies as context, but do not let "the protocol halts if operator X disappears" drive the access grade on its own; that belongs in dependencies. Source verification / contract verification on block explorers is assessed in the verifiability slice and is out of scope here — do NOT let "contract is unverified" drive the access grade.

Framing: the smart contracts are the access layer; frontends are UX. A permissionless contract is reachable by any client (SDK, third-party UI, aggregator, wallet integration). Frontend ToS, IP geo-blocking, and wallet screening are publisher policies on one specific client — they are reported as context but do NOT determine the grade. The grade hinges on (1) what the contract itself permits, and (2) whether the protocol is practically reachable without the official publisher's cooperation.

Meta-check before finalizing: if your verdict cites phrases from this prompt as evidence ("the protocol meets the 'credible alternatives' condition", "this fits the 'documented fallback' rule"), redo the verdict. The prompt describes the rubric; evidence must come from the protocol. A verdict should cite what the protocol does, not what the rubric says.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. Whitelist / allowlist modifiers in user-facing entry points. Grep for "onlyWhitelisted", "onlyRole", "allowlist", "isAccredited", "isKYCed". Note which functions are gated and who can add/remove from the list.
- A2. Off-chain operators in the admission path: keepers, sequencers, privileged relayers, oracle posters whose approval is required to admit a user action (not just to keep the protocol live). For each, identify whether the role is held by a single operator, a permissioned committee, or is permissionless. Enumerate per user-facing function class (deposit vs withdraw-request vs claim-finalized vs transfer) which ones require operator approval to be admitted, and which ones admit users unconditionally. A function whose placement is unconditional but whose downstream settlement depends on an operator is an admission-permissionless function — flag the liveness dependency as context and defer its grading weight to the dependencies slice.
- A3. Frontend restrictions on the official interface — record as context, not as a grade lever. Distinguish:
    - A3-passive: boilerplate ToS clauses (sanctions attestation, restricted-territory self-certification, VPN-circumvention prohibition, "comply with applicable law" eligibility, age of majority).
    - A3-active: runtime enforcement — IP-based geo-blocking, wallet-address screening against a sanctions oracle (Chainalysis, TRM, Elliptic), KYC wall, rendering-blocking jurisdiction banner.
  Record findings under the correct tier. Quote ToS text or banner text in evidence[].shows. These findings populate the headline and rationale but do NOT move the grade by themselves; the grade is set by A1, A2, and the A3b path check below.
- A3b. Independent access paths (the operative grade input). Enumerate paths that do not require the official publisher's cooperation:
    - Published SDK / library / CLI for direct contract interaction.
    - Third-party frontends operated by separate legal entities.
    - Wallet-integrated access (MetaMask Swaps, Safe apps, etc.).
    - DEX / lending / yield aggregators that route through the contracts.
  Record at least one concrete link per path that exists. The protocol does NOT have to self-document these — the test is existence, not UX cost. An A3b-i redistribution of the official UI bound by the same ToS does NOT count as an independent path.
- A4. Sanctions / compliance tooling at the contract level: does the protocol check addresses against OFAC lists or similar on-chain blocklists in the contract itself? (Frontend-only screening belongs in A3.)
- A5. Differentiate read access vs write access: many protocols are read-permissionless (anyone can view state) but write-gated (only certain addresses can deposit/borrow). Record both.
- A6. ToS / Legal links: locate them on the website and produce a VERBATIM quote of any jurisdictional, sanctions, or eligibility clause in evidence[].shows. If you cannot extract the clause text verbatim (SPA render failure, paywall, dead link, etc.), do NOT paraphrase or infer from general knowledge — record the ToS URL in unknowns[] with the reason extraction failed. Assertions about ToS content without a verbatim quote will be downweighted by reviewers.

Then write the steel-man section per Hard Rule 11.

Grade rules (admission-focused; liveness concerns belong in dependencies; source verification belongs in verifiability):
- green   = no contract-level whitelist/KYC on user entry/exit; no operator approval required to admit a user action; AND at least one independent A3b path exists (published SDK, third-party frontend, wallet integration, or aggregator routing). Frontend ToS posture and A3-active enforcement on the official UI do NOT block green when contracts are permissionless and an independent path exists — they are reported as context.
- orange  = contracts admit users unconditionally, BUT the protocol is operationally captured by the official publisher: no published SDK, no third-party frontend, no wallet integration, no aggregator routing. The contract is theoretically open but practically reachable only through the official UI. Also applies when admission requires approval from a permissioned committee that is governance-managed with a documented replacement procedure.
- red     = contract-level whitelist / KYC on user entry/exit, OR admission of a core user action requires approval from a single privileged operator or a small committee with no documented replacement procedure, OR enforces an on-chain blocklist updatable by a single party.
- unknown = checklist incomplete after checking the sources above.

Default-grade guidance: when contracts are fully permissionless AND any A3b independent path exists, the default grade is green regardless of frontend ToS or A3-active enforcement on the official UI. Frontend geo-blocking, sanctions-oracle wallet screening, and ToS sanctions clauses are publisher policies on one client and are reported in findings/headline as context, not as grade levers. To grade orange on operational-capture grounds, the auditor must affirmatively show that ALL independent paths are absent or also gated.

Guideline on committees: where admission depends on a multi-operator committee, the relevant axes are (a) set size, (b) whether replacement/rotation is governed on-chain, (c) whether the replacement procedure is publicly documented. A large set with on-chain governance replacement should not be graded as a single-party operator even if rotation is not instantaneous. A small set with informal replacement should be treated as a single-party operator.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "open-access",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Audit all 5 dimensions · one prompt Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v2
- protocol.name:              Balancer V2
- protocol.chains:            Ethereum, Polygon, Arbitrum, Base, xDai, Avalanche, Polygon zkEVM, Mode, Fraxtal
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/balancer/balancer-v2-monorepo/tree/master/audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: CONTROL

Evaluate who can change the protocol's rules, how fast, and how broadly.

(Step 0 capability probe and the off-chain-only fallback live in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[])

- **C1.** For each address you assess: who is the contract owner / admin / pendingAdmin / governor — read these via the block explorer's "Read Contract" tab OR `https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=owner` (BARE method names: `&method=owner`, `&method=admin`, `&method=pendingOwner`, `&method=governor`). For Safes use `/api/safe/owners?chainId=<id>&address=0x...`. When a protocol has multiple major versions deployed (v2/v3/v4), perform C1 reads on the NEWEST deployment separately — newer deployments often have weaker control surfaces than the legacy core.
- **C2.** Upgrade mechanism: transparent proxy / UUPS / Beacon / Diamond / immutable. Identify the proxy admin address. Check upgradeability of GOVERNANCE contracts too — a Governor / Aragon Voting / OZ Governor is often itself a proxy whose admin is the Timelock. Asymmetry: when fund-holding cores are immutable AND governance has no admin path that reaches them, an upgradable Governor/Timelock is T3-only and must NOT drag the verdict below green on that basis alone (see grade rules and the "immutable cores" caveat). Only call upgradability "mixed" if you can name a concrete function on the upgradable surface that reaches T1 or T2 on user funds.
- **C3.** EXECUTION PATH (enumerate every stage, in order, with delays in seconds). The operative path is usually a chain (voting → scheduler → timelock → executor; or governor → queue → execute; or Aragon Voting → DualGovernance → EmergencyProtectedTimelock → AdminExecutor). For each stage, record (a) the contract address, (b) the delay constant name + value in seconds, (c) the URL you read it from (block-explorer Read Contract OR `/api/contract/read?...&method=MIN_DELAY&block=<n>`). Do NOT stop at the first timelock-shaped contract — if its admin is itself called by another contract, keep walking. The grading delay is the SUM OF DELAYS ON THE UNCONTESTED FAST PATH (shortest time a proposal with no opposition can go from submission to executable). Dynamic / contested extensions (veto signaling, rage quit, escrow delay) are modifiers, not the basis — note them separately.
- **C4.** Enumerate EVERY multisig with reachable control — main proxy admin, emergency activation, emergency execution, reseal / pause, gate-seal committees, tiebreaker, per-module admins. For each Safe, fetch threshold + owners + version via `/api/safe/owners?chainId=<id>&address=0x...` (response includes raw eth_call data, so the URL is citable evidence). Enumerate ops/council/incentives multisigs even when off the upgrade path — record their scope so a reader can see they are NOT on the upgrade path. For each: (a) address, (b) threshold / total signers, (c) signer identities classified as insider (team, paid auditors under ongoing engagement, mandated service providers) vs non-insider (independent community members, unaffiliated researchers), (d) the specific power held (upgrade, pause, parameter, etc.).
- **C5.** On-chain governance: Governor / GovernorBravo / OZ Governor / Aragon Voting with token-weighted voting? Record proposal threshold, voting period, quorum, and the timelock delay between queue and execute. Every numeric constant must come from a Read Contract call you can link to, or be in unknowns[] with the C-code. If votingDelay / votingPeriod are denominated in BLOCKS, convert to seconds at the chain's CURRENT block time (Ethereum mainnet ≈ 12s post-Merge, not the 15s in older Compound/Bravo deployments) — cite both block count and converted seconds.
- **C6.** EMERGENCY POWERS: separate emergency-pause / guardian role with a different time cap or different actor than the main upgrade authority? Record it explicitly.
- **C7.** POWER TIER (blast radius). For each privileged path in C3–C6, classify the WORST thing that path can do, choosing the highest applicable tier. Cite the specific function name and any on-chain bound — tier claims without a named function are unsupported.
    - **T1 — FUND-CRITICAL**: replace implementation of contracts holding user funds; change AMM math / accounting / collateral logic; mint unbacked debt or shares; pause withdrawals; drain user-fund treasury; change oracle to attacker-controlled source; replace upgrade admin with EOA.
    - **T2 — ECONOMICALLY MATERIAL**: change fee parameters within bounded ranges; redirect protocol fees; add/remove markets / collateral types; bounded inflation or token mint within hard-capped schedule; spend protocol-owned (non-user) treasury.
    - **T3 — GOVERNANCE-INTERNAL**: change voting rules, quorum, voting period, proposal threshold; upgrade the Governor itself; rotate Timelock admin; mint governance tokens within a capped annual schedule.
    - **T4 — OPERATIONAL**: incentives distribution, grants, ENS / frontend canonicalization, deployment coordination, periphery router deprecation.
  The grade is set by the HIGHEST tier reachable on the uncontested fast path, not the median. State the tier and the binding function in the verdict.

### Read Contract discipline (applies to C3, C4, C5)

Every numeric constant cited (timelock delays, voting periods, multisig thresholds, quorum percentages) must come from EITHER (a) a block-explorer Read Contract URL, OR (b) a DeFiPunkd `/api/contract/read` or `/api/safe/owners` URL (preferred with `&block=<n>` for content-addressed evidence), OR (c) an unknowns[] entry with the C-code. Docs / blog posts are corroboration only — they cannot be the sole citation for a value that is also readable on-chain.

### Off-chain-only substitute hierarchy (when grading_basis="off-chain-only" — see preamble Rule 16)

When on-chain reads were genuinely unreachable this run, eligible off-chain substitutes in priority order:
1. Linked audit PDFs (admin roles, multisig members, timelock delays usually enumerated).
2. Governance forum posts that quote constants from a successful on-chain proposal (cite post URL + linked execution-tx URL).
3. Official protocol docs pages with named addresses and roles (must be on a domain owned by the protocol).
4. GitHub README / SECURITY.md / governance/*.md at a pinned commit SHA.

Forbidden substitutes: third-party blog posts, X / Twitter threads, search-result snippets, model memory. Required degradation: any C-code citing a numeric constant from docs/forum/audit prose ONLY must also carry an `unknowns[]` entry with `-offchain` suffix noting "value not re-read on-chain in this run; corroboration only".

### Grade rules (apply the timelock bar conditional on the highest C7 tier reachable on the fast path)

Security Council standard (used below): a multisig qualifies as "Security Council" only if ALL of: ≥7 signers, ≥51% threshold, ≥50% non-insider signers, every signer publicly announced. Failing any criterion = NOT a Security Council, regardless of signer reputation.

- **green**: highest reachable tier is T3 or T4 regardless of timelock; OR T2 reachable with uncontested-fast-path delay ≥7 days; OR T1 reachable only via immutable contracts (T1 is unreachable); OR T1 reachable with uncontested-fast-path delay ≥7 days combined with a Security Council multisig; OR T1 reachable with uncontested-fast-path delay ≥7 days through active on-chain governance with broad token distribution.
- **orange**: T2 reachable with uncontested-fast-path delay >0 but <7 days; OR T1 reachable with uncontested-fast-path delay >0 but <7 days; OR a multisig failing one or more Security Council criteria sits on a T1/T2 path; OR unclear upgrade authority on a T1/T2 path; OR governance with very short timelock or low quorum on a T1/T2 path.
- **red**: T1 reachable with no timelock by a single EOA or 2-of-3 multisig; OR a T1 upgrade admin that is not a smart contract you can audit.
- **unknown**: completed the checklist but still cannot determine the upgrade authority OR cannot classify the highest tier reachable on the main contracts.

Tiering caveats:
- "Bounded" must be enforced ON-CHAIN to count as T2. A function that sets fees with no upper-bound check is T1 — cite the bound check.
- Recurring T2 economic extraction (e.g. fee redirect with no rate limit) approaches T1 over time. A single proposal that can permanently redirect all future revenue is T1.
- T3 assumes the governance contract cannot itself authorize a T1/T2 action without going through the same timelock. If governance can self-upgrade to bypass the timelock, T3 collapses into T1.
- Do not downgrade tier by hand-waving ("realistically governance would never…"). Tier on what the contract permits, not what feels likely.

Notes:
- **Dynamic / dual-governance timelocks** (Lido, Compound escrow veto): the rubric grades on the uncontested path because that is the path most upgrades take. A dynamic extension that fires only under stake-weighted opposition is a real protection — name it in the green steel-man, but it does not lift an orange fast path into green; state the tension in the verdict.
- **Immutable cores with upgradable governance** (Uniswap-style): if fund-holding contracts are immutable and have no admin-reachable function moving / freezing / re-routing user funds, the highest reachable tier on the upgrade path is T3 — green regardless of timelock. Don't grade this orange just because the Governor is a proxy — that's a C2 fact, not a downgrade. Downgrade only applies if you can cite a concrete function on the upgradable surface that reaches T1 or T2 (privileged hook, upgradable factory controlling fund-routing, fee-switch redirecting protocol revenue without bound).
- **The 7-day bar** reflects the exit-window standard — users need notice after a queued upgrade to withdraw if they disagree. The ability-to-exit slice grades the exit side; this slice grades the delay side; both must hold for users to actually benefit from the delay.

---

### Slice: ABILITY-TO-EXIT

Evaluate whether users can withdraw their funds on their own terms, even under adversarial admin conditions.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- E1. Enumerate every user-facing exit function in the main contracts: withdraw, redeem, burn, requestWithdrawal, claim, exit, etc. List them by name. Do NOT treat the contract as a monolith.
- E2. For EACH exit function in E1: identify its access modifiers and any pause guards (e.g. _checkResumed, whenNotPaused, onlyRole). Functions that gate REQUEST PLACEMENT often differ from functions that CLAIM FINALIZED FUNDS — check both separately.
- E3. For each pause guard: identify the role holder (which address holds PAUSE_ROLE / GUARDIAN / etc.) and the maximum pause duration. Specifically check whether PAUSE_INFINITELY (or equivalent uncapped pause) is callable, and which actor can call it (single multisig vs governance vote). For role-holder reads use https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=hasRole&args=0x...,0x... or &method=getRoleAdmin&args=0x.... For "is currently paused" checks use &method=paused or &method=isPaused&args=<resume-code>. Use the BARE method name (no parens). Cite the URL with &block=<n> in evidence[].
- E4. EMERGENCY vs GOVERNANCE pause distinction: many protocols have a fast-acting emergency pause capped at N days and a slow governance pause that can be indefinite. Record both paths separately if present, with their time caps and actor classes.
- E5. Queued redemption: documented maximum queue duration, daily withdrawal caps, whether the queue itself is pausable.
- E6. Forced-exit / escape-hatch / permissionless emergency-exit mechanism for adversarial-admin scenarios.
- E7. Frontend dependency: confirm exit functions are directly callable on-chain (e.g. via Etherscan write tab or a generic wallet) without the project's frontend.

Then write the steel-man section per Hard Rule 11. Common red-vs-orange tension on this slice: indefinite pause exists (suggests red) BUT the realistic emergency path is time-capped AND claims of already-finalized exits are not pause-gated (suggests orange). Resolve this by stating who can do what for how long, not by stopping at the worst-case sentence.

Grade rules:
- green   = permissionless exit; pause is either absent, narrowly scoped to clearly-described emergencies with auto-expiry, or capped at ≤7 days; no frontend dependency for exit; claims of already-finalized exits are not pause-gated under any path.
- orange  = pausable with broad scope OR indefinite pause is reachable only through governance vote (not unilateral admin action), OR queued redemption with documented max > 7 days, OR claims-of-finalized are exempt but new-request placement can be paused indefinitely by governance.
- red     = exit requires admin signature, OR ANY actor (including governance) can pause CLAIMS of finalized exits indefinitely, OR there is no on-chain exit function at all (purely custodial), OR pause is held by a single EOA / 2-of-3 multisig with no time cap.
- unknown = checklist incomplete after checking the sources above.

---

### Slice: AUTONOMY

Evaluate this protocol's autonomy: can a failure of anything outside its own contracts cause theft or loss of user principal, loss of unclaimed yield, or materially change the protocol's expected performance? "Autonomy" is not "has zero external touchpoints" — it is "external failures are either survivable or recoverable without user loss". This slice adapts DefiScan v1's Autonomy dimension; dependencies are one of several criteria, not the whole frame.

GUARDRAILS (read before grading):
- Category alone (Liquid Staking, Bridge, RWA Lending, Restaking, …) does NOT force a grade. A category is a hint about where to look; the grade must come from the concrete A1–A9 findings below.
- Base-chain consensus (Ethereum PoS, the chain's validator set, the canonical Deposit Contract at 0x00000000219ab540356cBB839Cbe05303d7705Fa) is the SUBSTRATE, not a dependency, for any protocol deployed on that chain. Do not list "depends on Ethereum" as a finding.
- Oracles or other integrations used by DOWNSTREAM protocols that happen to read this protocol's token (e.g. Chainlink stETH/USD consumed by Aave) are NOT this protocol's dependencies. Count only what THIS protocol's contracts call or trust on-chain. If a feed or contract appears in the protocol's docs only as reference material for third-party integrators, EXCLUDE IT ENTIRELY — do not log it as a finding even with a "peripheral" or "referenced only" caveat.
- "Upgradeable admin can change things" belongs to the CONTROL slice; only count it here when the upgrade surface lets governance silently swap an external dependency (see A9).
- Underlying-asset risk in opt-in, isolated markets is NOT autonomy-red on its own. When a protocol wraps third-party yield-bearing assets (LSTs, LRTs, ERC-4626 vaults, lending receipts, restaked tokens) into per-market silos that users explicitly choose, a failure of one underlying does NOT propagate to other markets and is risk the user opted into per-market. Record it under A4 with depth and propagation scope, but do not let it alone drive a red verdict — red requires an external dependency that cross-cuts the protocol or that the user did not opt into at deposit time. A failure mode that is "if the LRT you deposited is hacked, your principal in that LRT-backed market is impaired" is the underlying's autonomy story, not this protocol's; grade it on whether THIS protocol introduces additional dependencies on top.
- Sub-module enumeration is mandatory before grading. If the protocol ships distinct product lines or modules (e.g., a v2 core AMM plus a newer perps/funding-rate module, a lending pool plus a separate vault layer, an L1 core plus a cross-chain extension), enumerate each in findings and grade against the WORST module weighted by its share of TVS or its blast radius. A green core does not rescue an orange/red sub-module; conversely, a small red sub-module with capped TVS may bound the overall grade to orange. Name each module by its on-chain factory or router address. If you do not know whether a module exists, that is an unknowns[] entry, not silence. EXPLICIT TVS WEIGHTING: in the verdict, state each module's approximate share of total protocol TVS (use "~X%" if exact figures unavailable; check DeFiLlama or block-explorer balances on the module's main contract) and how that share informs the weighted grade. Format: "Module A holds ~X% of TVS (grade: <g>); Module B holds ~Y% (grade: <g>); weighted overall = <grade> because <reason>." If a red sub-module holds <5% of total TVS and is capped, the overall grade may be orange; if it holds >25%, the overall grade is red. Do not let qualitative reasoning substitute for the percentages — write the numbers.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. External contract calls. Enumerate every external contract the core contracts call or read from (oracles, price feeds, AMM pools, lending pools, staking/deposit contracts, yield wrappers). For each, identify the address, the provider, and what user-facing function of this protocol would break or mis-price if that external contract paused, mis-reported, or behaved adversarially. Grep for "oracle", "aggregator", "getPrice", "latestAnswer", "chainlink", "pyth", "redstone" as a starting point. To verify an oracle is live and what it currently reports, hit https://defipunkd.com/api/contract/read?chainId=<id>&address=<oracle>&method=latestAnswer (or &method=latestRoundData, &method=getPrice, &method=description; use the BARE method name without parens) and cite the URL — the response includes blockNumber/blockHash and rawReturnData, which is stronger evidence than a docs page about which feed "should" be used.
- A2. Off-chain actor committees reporting INTO the protocol. Oracle committees, guardian multisigs, DAO-selected validator sets acting as protocol reporters, exit-bus signers, fraud-proof challengers. For each, record committee size, quorum, who picks the members, and what mis-reporting could do (mint, burn, finalize withdrawal, freeze). Distinguish this from governance admins (control slice). NOTE ON STAKING PROTOCOLS: validator slashing and node-operator misbehavior are properties of the base-chain substrate, NOT external dependencies, when the operator set is diversified enough that a coordinated failure caps at <5% principal loss. Count validator/operator risk under A2 ONLY if the operator set is small, non-diversified, or lacks bonding / slashing-insurance / diversification mitigations. A curated set of 30+ independent operators with documented diversification falls under the mitigated path; a 3-operator LST does not. Do not cite the protocol's own risk disclosure as evidence that operator failure = principal loss unless you also check the diversification and bond mitigations.
- A3. Bridge / cross-chain messaging dependencies. Only count bridges that carry material TVL or are required for a core user flow. For each, name the bridge operator (canonical L1↔L2, LayerZero, Wormhole, Axelar, custom multisig), the trust model (canonical, optimistic, light-client, guardian set), and what fraction of TVL or users ride on it. "wstETH exists on 15 chains" is not a finding unless material TVL sits there. Before listing any non-primary chain deployment as a dependency, verify it is still operational as of analysis_date — retired or sunset deployments (e.g., Lido-on-Terra, Lido-on-Solana) belong in unknowns[] or should be omitted, never cited as a current dependency.
- A4. Nested collateral / restaking chains. For restaking / LRT / receipt-of-receipt designs: record the depth of the collateral chain, every actor with slashing or freezing power at each level, and whether a failure N levels deep propagates to user principal here.
- A5. Fork lineage (silent check). If DeFiLlama's forkedFrom is non-empty, record it as one finding and move on. If empty, do NOT add a placeholder finding; it adds noise.
- A6. Fallback mechanisms and circuit breakers. What catches an external failure? Sanity-check contracts on oracle reports, rebase bounds, pause paths triggered by bad prices, second-opinion oracles, max-per-block throughput caps, withdrawal queues that absorb bad reports. Record which A1–A4 risks are mitigated by which fallback, and which are unmitigated. For EACH fallback you cite, state its activation status explicitly: (i) LIVE and enforcing on-chain today, (ii) DEPLOYED but not yet wired / activated (e.g., interface exists but the address is zero or the role is unassigned), or (iii) DOCUMENTED / PROPOSED only (forum post, LIP draft, audit pending). Only (i) counts as mitigation for the grade. A fallback in state (ii) or (iii) should be noted but must not reduce the risk in your steel-man or verdict. If you cannot determine activation status, add an unknowns[] entry rather than assume it is live.
- A7. Sequencer / L1-liveness dependency BEYOND the base-chain substrate. SCOPE — sequencer risk only counts here when the protocol IS its own L2/L3 appchain or app-rollup, where the sequencer is part of the protocol's own stack and a freeze is a protocol-level outage. A protocol permissionlessly deployed on a third-party L2 inherits that L2's sequencer as substrate, not an A7 dependency. Record the sequencer/DA trust model when A7 applies.
- A8. Keeper / relayer / off-chain bot liveness. Protocols that need permissionless-but-necessary off-chain actors (liquidation bots, auto-compounders, deposit relayers, intent solvers). Record whether the role is permissionless, what degrades if nobody runs it (yield paused, bad debt accumulates, positions go stale), and whether the failure mode is graceful or catastrophic.
- A9. Governance-mutable dependency surface. Can an admin or DAO action silently INTRODUCE a NEW EXTERNAL dependency — swap the oracle address to a different provider, register a new staking module that calls an untrusted contract, add a new bridge, route SY through a new external vault — without an exit window for users? Check the upgrade / router / module-registry contracts. Answer: which EXTERNAL dependencies are governance-mutable, who holds that power, and whether there is a timelock or exit window. SCOPE LIMIT — read carefully: A9 is about the *external dependency surface*, not the upgrade surface in general. "The proxyAdmin / EOA can upgrade the router implementation to arbitrary bytecode" is a CONTROL-slice finding (admin can rug), NOT an autonomy-A9 finding. A9 fires only when the upgrade specifically swaps out or adds a contract that THIS protocol calls or trusts (e.g., changing the oracle address from Chainlink to a malicious feed, registering a new SY adapter that points to a third-party vault, redirecting a bridge endpoint). If the only finding is "admin can change implementation," do not log it under A9 and do not let it drive the autonomy grade — note it under control instead. The autonomy-relevant version of the same upgrade key is "admin can swap [specific external address X] without timelock"; that requires identifying the specific external dependency that becomes mutable.

STEEL-MAN (per Hard Rule 13): write one-sentence strongest arguments for red, orange, and green using the A1–A9 findings.

IMPACTED TVS ESTIMATE: the headline MUST include a rough impacted-TVS figure — the fraction of protocol TVS that could be lost or frozen if the worst-unmitigated dependency you identified failed. Use "~X%" if exact numbers are unavailable, "<1%" for de minimis, "unclear" only if A1–A9 left the question genuinely open (in which case grade=unknown is usually correct). Do NOT substitute qualitative phrases like "significant" — give a number or bracket.

GRADE ANCHORS (mapped to DefiScan v1 stages):
- green = Stage 2 equivalent. Failure of any external dependency cannot cause loss of user principal or unclaimed yield. Either there are no material external dependencies, or every critical one has a documented fallback (A6) that keeps users whole. Governance cannot silently introduce new dependencies without an exit window (A9). Impacted TVS under any single-dependency failure: effectively 0.
- orange = Stage 1 equivalent. Failure of some external dependency can cause loss of unclaimed yield, or can materially change expected performance (pause withdrawals, freeze positions, degrade price quality), but cannot cause loss of principal. Committee-based oracles with sanity checks, canonical-only bridges, fallback paths that exist but are incomplete, or governance-mutable dependencies protected by a ≥7-day timelock. Impacted TVS is bounded and recoverable.
- red = Stage 0 equivalent. Failure of an external dependency CAN cause theft or loss of principal. Examples: single-provider oracle with no sanity check or fallback, material TVL on a non-canonical bridge with a guardian multisig, governance can hot-swap oracles or add staking modules with no timelock or exit window, unmitigated keeper-liveness dependency where positions become insolvent if bots stop. Impacted TVS is material.
- unknown = checklist incomplete after inspecting source + verified contracts. Prefer unknown over guessing when A1/A6 could not be reconstructed. RESERVE unknown for cases where the CORE ARCHITECTURE itself is unverifiable — not for cases where you merely cannot enumerate every per-market dependency in a multi-market protocol. If the core router/factory/oracle architecture is verifiable on-chain and you can determine whether the core requires external dependencies, grade the architecture even when an exhaustive per-market external-dependency census is infeasible. Acknowledge the per-market gap in unknowns[] but still issue a grade. Refusing to grade a multi-market protocol because you cannot list every SY/vault/market is over-use of unknown; grade the architecture and say so.

PROMPT-META CHECK (per Hard Rule 17): before finalizing, verify the verdict cites concrete contract addresses, docs, or code — not the rubric itself. If your verdict says "the protocol belongs to a category the rubric marks red", rewrite it with the A1–A9 finding that actually justifies the grade, or drop to grade=unknown.

---

### Slice: OPEN-ACCESS

Evaluate who is allowed to use the protocol and whether any of that permission is granted off-chain.

Scope: this slice is about ADMISSION — who can enter, exit, or transact. Operator LIVENESS (what breaks if keepers/oracles go offline) is assessed in the dependencies slice and is out of scope for the grade here. You may note operator dependencies as context, but do not let "the protocol halts if operator X disappears" drive the access grade on its own; that belongs in dependencies. Source verification / contract verification on block explorers is assessed in the verifiability slice and is out of scope here — do NOT let "contract is unverified" drive the access grade.

Framing: the smart contracts are the access layer; frontends are UX. A permissionless contract is reachable by any client (SDK, third-party UI, aggregator, wallet integration). Frontend ToS, IP geo-blocking, and wallet screening are publisher policies on one specific client — they are reported as context but do NOT determine the grade. The grade hinges on (1) what the contract itself permits, and (2) whether the protocol is practically reachable without the official publisher's cooperation.

Meta-check before finalizing: if your verdict cites phrases from this prompt as evidence ("the protocol meets the 'credible alternatives' condition", "this fits the 'documented fallback' rule"), redo the verdict. The prompt describes the rubric; evidence must come from the protocol. A verdict should cite what the protocol does, not what the rubric says.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. Whitelist / allowlist modifiers in user-facing entry points. Grep for "onlyWhitelisted", "onlyRole", "allowlist", "isAccredited", "isKYCed". Note which functions are gated and who can add/remove from the list.
- A2. Off-chain operators in the admission path: keepers, sequencers, privileged relayers, oracle posters whose approval is required to admit a user action (not just to keep the protocol live). For each, identify whether the role is held by a single operator, a permissioned committee, or is permissionless. Enumerate per user-facing function class (deposit vs withdraw-request vs claim-finalized vs transfer) which ones require operator approval to be admitted, and which ones admit users unconditionally. A function whose placement is unconditional but whose downstream settlement depends on an operator is an admission-permissionless function — flag the liveness dependency as context and defer its grading weight to the dependencies slice.
- A3. Frontend restrictions on the official interface — record as context, not as a grade lever. Distinguish:
    - A3-passive: boilerplate ToS clauses (sanctions attestation, restricted-territory self-certification, VPN-circumvention prohibition, "comply with applicable law" eligibility, age of majority).
    - A3-active: runtime enforcement — IP-based geo-blocking, wallet-address screening against a sanctions oracle (Chainalysis, TRM, Elliptic), KYC wall, rendering-blocking jurisdiction banner.
  Record findings under the correct tier. Quote ToS text or banner text in evidence[].shows. These findings populate the headline and rationale but do NOT move the grade by themselves; the grade is set by A1, A2, and the A3b path check below.
- A3b. Independent access paths (the operative grade input). Enumerate paths that do not require the official publisher's cooperation:
    - Published SDK / library / CLI for direct contract interaction.
    - Third-party frontends operated by separate legal entities.
    - Wallet-integrated access (MetaMask Swaps, Safe apps, etc.).
    - DEX / lending / yield aggregators that route through the contracts.
  Record at least one concrete link per path that exists. The protocol does NOT have to self-document these — the test is existence, not UX cost. An A3b-i redistribution of the official UI bound by the same ToS does NOT count as an independent path.
- A4. Sanctions / compliance tooling at the contract level: does the protocol check addresses against OFAC lists or similar on-chain blocklists in the contract itself? (Frontend-only screening belongs in A3.)
- A5. Differentiate read access vs write access: many protocols are read-permissionless (anyone can view state) but write-gated (only certain addresses can deposit/borrow). Record both.
- A6. ToS / Legal links: locate them on the website and produce a VERBATIM quote of any jurisdictional, sanctions, or eligibility clause in evidence[].shows. If you cannot extract the clause text verbatim (SPA render failure, paywall, dead link, etc.), do NOT paraphrase or infer from general knowledge — record the ToS URL in unknowns[] with the reason extraction failed. Assertions about ToS content without a verbatim quote will be downweighted by reviewers.

Then write the steel-man section per Hard Rule 11.

Grade rules (admission-focused; liveness concerns belong in dependencies; source verification belongs in verifiability):
- green   = no contract-level whitelist/KYC on user entry/exit; no operator approval required to admit a user action; AND at least one independent A3b path exists (published SDK, third-party frontend, wallet integration, or aggregator routing). Frontend ToS posture and A3-active enforcement on the official UI do NOT block green when contracts are permissionless and an independent path exists — they are reported as context.
- orange  = contracts admit users unconditionally, BUT the protocol is operationally captured by the official publisher: no published SDK, no third-party frontend, no wallet integration, no aggregator routing. The contract is theoretically open but practically reachable only through the official UI. Also applies when admission requires approval from a permissioned committee that is governance-managed with a documented replacement procedure.
- red     = contract-level whitelist / KYC on user entry/exit, OR admission of a core user action requires approval from a single privileged operator or a small committee with no documented replacement procedure, OR enforces an on-chain blocklist updatable by a single party.
- unknown = checklist incomplete after checking the sources above.

Default-grade guidance: when contracts are fully permissionless AND any A3b independent path exists, the default grade is green regardless of frontend ToS or A3-active enforcement on the official UI. Frontend geo-blocking, sanctions-oracle wallet screening, and ToS sanctions clauses are publisher policies on one client and are reported in findings/headline as context, not as grade levers. To grade orange on operational-capture grounds, the auditor must affirmatively show that ALL independent paths are absent or also gated.

Guideline on committees: where admission depends on a multi-operator committee, the relevant axes are (a) set size, (b) whether replacement/rotation is governed on-chain, (c) whether the replacement procedure is publicly documented. A large set with on-chain governance replacement should not be graded as a single-party operator even if rotation is not instantaneous. A small set with informal replacement should be treated as a single-party operator.

---

### Slice: VERIFIABILITY

Evaluate whether an outsider can independently confirm what the deployed code does.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- V1. For each address you assess: is the bytecode verified on the chain's block explorer? Record the "Contract Source Code Verified" indicator. https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x... returns this as a top-level "verified" boolean plus "abiSource" ("etherscan" / "sourcify") and an inline ABI — useful when the explorer page is rate-limited. If the contract is a proxy, verify BOTH the proxy contract AND the current implementation contract. The same /api/contract/abi response auto-resolves proxies and includes a "proxy.implementation" address when present, so one call covers V1 + V6 in one shot. An explorer "Similar Match" on a well-known proxy pattern (Aragon AppProxyUpgradeable, ERC1967Proxy, OssifiableProxy, OZ TransparentUpgradeableProxy) is expected for that pattern and does NOT count as a verification gap on its own — what matters is that the implementation is independently verified.
- V2. Source-to-repo correspondence: for each verified contract, attempt to find a matching commit in the linked GitHub repos. Record evidence[].commit on a match. Independent compile/bytecode-match is NOT required for green — a recognized public repo whose structure and file contents correspond to the explorer-visible source is sufficient. If you did not pin a commit SHA or run a bytecode diff, record that plainly in unknowns[] and proceed; it is a scope limit, not a downgrade signal.
- V3. Audit coverage: for each URL in protocol.audit_links, open it and record: auditor name, audit date, the specific contracts / commit in scope. Flag audits that predate the current deployment by >6 months without a follow-up review.
- V4. Auditor recognition: the following firms are broadly recognized in Solidity: Trail of Bits, Zellic, Spearbit, OpenZeppelin, ConsenSys Diligence, Certora (formal verification), Quantstamp, Halborn, Peckshield, Sigma Prime, ChainSecurity, Ackee Blockchain, MixBytes, Statemind. Unknown firms are orange-at-best for any green-grade claim. Name the firm explicitly in evidence[].
- V5. Post-audit drift: compare the most recent in-scope audit(s) against the currently-deployed source, weighted by what each contract does and by what the changes actually contain. SCOPE — drift only downgrades the grade when ALL of the following hold: (i) the drifting contract is fund-custody / settlement / accounting-critical (NOT a peripheral router, lens, quoter, or pure-view contract that holds no balances); (ii) the changes are material — new functions, modified access control, modified accounting, modified fund flow — and not refactors / struct relocations / import reorgs / build or CI fixes / formatting; (iii) no later audit, fix-audit, or differential audit from a recognized firm covers the changed files (audits often pin a pre-fix commit while a follow-up reviews the delta — match by file scope, not by commit-hash equality). When you cite drift as a downgrade reason, name the specific behavior change (function added, role granted, accounting formula altered) — "N commits ahead" or "+X/-Y LOC" alone is not evidence of material drift. If you have not sampled the diff content (e.g. via the GitHub compare view or the top commits in the window), record drift as an unknowns[] entry rather than auto-downgrading; commit-count and LOC are starting signals, not findings.
- V6. Implementation vs proxy: a verified proxy with an unverified implementation is effectively unverified. State whether the implementation is verified separately.

EVIDENCE DISCIPLINE (read before writing findings[]):
- Do not assert a specific deploy-commit SHA, bytecode equivalence, or "identical to audited commit" unless you actually fetched the artifact that shows it (e.g., a deployed-addresses JSON you opened, an explorer page you read). Inferred or plausible matches belong in unknowns[], never in findings[] or evidence[].
- Evidence[] entries must correspond to pages/files you actually retrieved this run. A URL you did not open is not evidence.

Then write the steel-man section per Hard Rule 11.

Grade rules:
- green   = deployed bytecode verified on the explorer (proxy AND implementation if proxied; "Similar Match" on a standard proxy pattern is fine per V1), a public source repo exists whose contents correspond to the explorer-visible source, AND ≥1 audit from a recognized firm covering the currently-deployed contracts (≤6 months of drift OR drift was re-audited). A missing local compile-match is not a downgrade — record it in unknowns[] and still grade green if the other conditions hold.
- orange  = verified but with visible drift from the public repo, OR audit scope is stale relative to deployment, OR only minor / unknown-firm audits exist, OR only some of the main contracts are verified, OR proxy verified but implementation only partially verified.
- red     = unverified bytecode (or verified proxy with unverified implementation), OR no audit in protocol.audit_links, OR no public repo.
- unknown = reserved for when the protocol's verifiability posture genuinely cannot be assessed (e.g., explorer and repo both inaccessible for this protocol). Do NOT use unknown merely because you, the analyst, could not run a particular check such as a bytecode diff — that goes in unknowns[] while the grade is still assigned from the evidence you do have.

---

### JSON output contract

Return exactly one JSON array inside a single ```json fenced block. The array MUST contain exactly five objects, one per risk slice, in this exact order: "control", "ability-to-exit", "autonomy", "open-access", "verifiability". Do not include the discovery slice.

Each object has the same shape as a normal slice submission:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "<one of: control | ability-to-exit | autonomy | open-access | verifiability>",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- Produce one complete object for each of these slices only: control, ability-to-exit, autonomy, open-access, verifiability.
- Reuse the same model, chat_url, snapshot_generated_at, prompt_version, analysis_date, and slug values across all five objects.
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches that object's slice checklist prefix verbatim (C1, E2, AU3, A3b, V4a, …); unknowns[] entries are checklist-coded ("C3: …").
- Wrap the array in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Verifiability tentative

Open source + 3 audits

Protocol publishes a GitHub repository and has at least one audit on record. This is a coarse Phase-0 signal only: auditor reputation, scope, and post-audit review coverage are not yet weighted.

Run your own prompt Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v2
- protocol.name:              Balancer V2
- protocol.chains:            Ethereum, Polygon, Arbitrum, Base, xDai, Avalanche, Polygon zkEVM, Mode, Fraxtal
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/balancer/balancer-v2-monorepo/tree/master/audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: VERIFIABILITY

Evaluate whether an outsider can independently confirm what the deployed code does.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- V1. For each address you assess: is the bytecode verified on the chain's block explorer? Record the "Contract Source Code Verified" indicator. https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x... returns this as a top-level "verified" boolean plus "abiSource" ("etherscan" / "sourcify") and an inline ABI — useful when the explorer page is rate-limited. If the contract is a proxy, verify BOTH the proxy contract AND the current implementation contract. The same /api/contract/abi response auto-resolves proxies and includes a "proxy.implementation" address when present, so one call covers V1 + V6 in one shot. An explorer "Similar Match" on a well-known proxy pattern (Aragon AppProxyUpgradeable, ERC1967Proxy, OssifiableProxy, OZ TransparentUpgradeableProxy) is expected for that pattern and does NOT count as a verification gap on its own — what matters is that the implementation is independently verified.
- V2. Source-to-repo correspondence: for each verified contract, attempt to find a matching commit in the linked GitHub repos. Record evidence[].commit on a match. Independent compile/bytecode-match is NOT required for green — a recognized public repo whose structure and file contents correspond to the explorer-visible source is sufficient. If you did not pin a commit SHA or run a bytecode diff, record that plainly in unknowns[] and proceed; it is a scope limit, not a downgrade signal.
- V3. Audit coverage: for each URL in protocol.audit_links, open it and record: auditor name, audit date, the specific contracts / commit in scope. Flag audits that predate the current deployment by >6 months without a follow-up review.
- V4. Auditor recognition: the following firms are broadly recognized in Solidity: Trail of Bits, Zellic, Spearbit, OpenZeppelin, ConsenSys Diligence, Certora (formal verification), Quantstamp, Halborn, Peckshield, Sigma Prime, ChainSecurity, Ackee Blockchain, MixBytes, Statemind. Unknown firms are orange-at-best for any green-grade claim. Name the firm explicitly in evidence[].
- V5. Post-audit drift: compare the most recent in-scope audit(s) against the currently-deployed source, weighted by what each contract does and by what the changes actually contain. SCOPE — drift only downgrades the grade when ALL of the following hold: (i) the drifting contract is fund-custody / settlement / accounting-critical (NOT a peripheral router, lens, quoter, or pure-view contract that holds no balances); (ii) the changes are material — new functions, modified access control, modified accounting, modified fund flow — and not refactors / struct relocations / import reorgs / build or CI fixes / formatting; (iii) no later audit, fix-audit, or differential audit from a recognized firm covers the changed files (audits often pin a pre-fix commit while a follow-up reviews the delta — match by file scope, not by commit-hash equality). When you cite drift as a downgrade reason, name the specific behavior change (function added, role granted, accounting formula altered) — "N commits ahead" or "+X/-Y LOC" alone is not evidence of material drift. If you have not sampled the diff content (e.g. via the GitHub compare view or the top commits in the window), record drift as an unknowns[] entry rather than auto-downgrading; commit-count and LOC are starting signals, not findings.
- V6. Implementation vs proxy: a verified proxy with an unverified implementation is effectively unverified. State whether the implementation is verified separately.

EVIDENCE DISCIPLINE (read before writing findings[]):
- Do not assert a specific deploy-commit SHA, bytecode equivalence, or "identical to audited commit" unless you actually fetched the artifact that shows it (e.g., a deployed-addresses JSON you opened, an explorer page you read). Inferred or plausible matches belong in unknowns[], never in findings[] or evidence[].
- Evidence[] entries must correspond to pages/files you actually retrieved this run. A URL you did not open is not evidence.

Then write the steel-man section per Hard Rule 11.

Grade rules:
- green   = deployed bytecode verified on the explorer (proxy AND implementation if proxied; "Similar Match" on a standard proxy pattern is fine per V1), a public source repo exists whose contents correspond to the explorer-visible source, AND ≥1 audit from a recognized firm covering the currently-deployed contracts (≤6 months of drift OR drift was re-audited). A missing local compile-match is not a downgrade — record it in unknowns[] and still grade green if the other conditions hold.
- orange  = verified but with visible drift from the public repo, OR audit scope is stale relative to deployment, OR only minor / unknown-firm audits exist, OR only some of the main contracts are verified, OR proxy verified but implementation only partially verified.
- red     = unverified bytecode (or verified proxy with unverified implementation), OR no audit in protocol.audit_links, OR no public repo.
- unknown = reserved for when the protocol's verifiability posture genuinely cannot be assessed (e.g., explorer and repo both inaccessible for this protocol). Do NOT use unknown merely because you, the analyst, could not run a particular check such as a bytecode diff — that goes in unknowns[] while the grade is still assigned from the evidence you do have.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "verifiability",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

4 dimensions not yet assessed (Control, Ability to exit, Autonomy, Open Access)

Control unknown Unverified

Not yet assessed

Who holds admin privileges, how contracts can be upgraded, and how quickly. No automated heuristic grades this at Phase 0; a real assessment arrives when onchain discovery reads roles, owners, and timelocks.

No model has graded this dimension yet. Run the slice prompt through any LLM and submit the JSON — once ≥3 independent runs agree, the quorum bot merges the verdict here.

Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v2
- protocol.name:              Balancer V2
- protocol.chains:            Ethereum, Polygon, Arbitrum, Base, xDai, Avalanche, Polygon zkEVM, Mode, Fraxtal
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/balancer/balancer-v2-monorepo/tree/master/audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: CONTROL

Evaluate who can change the protocol's rules, how fast, and how broadly.

(Step 0 capability probe and the off-chain-only fallback live in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[])

- **C1.** For each address you assess: who is the contract owner / admin / pendingAdmin / governor — read these via the block explorer's "Read Contract" tab OR `https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=owner` (BARE method names: `&method=owner`, `&method=admin`, `&method=pendingOwner`, `&method=governor`). For Safes use `/api/safe/owners?chainId=<id>&address=0x...`. When a protocol has multiple major versions deployed (v2/v3/v4), perform C1 reads on the NEWEST deployment separately — newer deployments often have weaker control surfaces than the legacy core.
- **C2.** Upgrade mechanism: transparent proxy / UUPS / Beacon / Diamond / immutable. Identify the proxy admin address. Check upgradeability of GOVERNANCE contracts too — a Governor / Aragon Voting / OZ Governor is often itself a proxy whose admin is the Timelock. Asymmetry: when fund-holding cores are immutable AND governance has no admin path that reaches them, an upgradable Governor/Timelock is T3-only and must NOT drag the verdict below green on that basis alone (see grade rules and the "immutable cores" caveat). Only call upgradability "mixed" if you can name a concrete function on the upgradable surface that reaches T1 or T2 on user funds.
- **C3.** EXECUTION PATH (enumerate every stage, in order, with delays in seconds). The operative path is usually a chain (voting → scheduler → timelock → executor; or governor → queue → execute; or Aragon Voting → DualGovernance → EmergencyProtectedTimelock → AdminExecutor). For each stage, record (a) the contract address, (b) the delay constant name + value in seconds, (c) the URL you read it from (block-explorer Read Contract OR `/api/contract/read?...&method=MIN_DELAY&block=<n>`). Do NOT stop at the first timelock-shaped contract — if its admin is itself called by another contract, keep walking. The grading delay is the SUM OF DELAYS ON THE UNCONTESTED FAST PATH (shortest time a proposal with no opposition can go from submission to executable). Dynamic / contested extensions (veto signaling, rage quit, escrow delay) are modifiers, not the basis — note them separately.
- **C4.** Enumerate EVERY multisig with reachable control — main proxy admin, emergency activation, emergency execution, reseal / pause, gate-seal committees, tiebreaker, per-module admins. For each Safe, fetch threshold + owners + version via `/api/safe/owners?chainId=<id>&address=0x...` (response includes raw eth_call data, so the URL is citable evidence). Enumerate ops/council/incentives multisigs even when off the upgrade path — record their scope so a reader can see they are NOT on the upgrade path. For each: (a) address, (b) threshold / total signers, (c) signer identities classified as insider (team, paid auditors under ongoing engagement, mandated service providers) vs non-insider (independent community members, unaffiliated researchers), (d) the specific power held (upgrade, pause, parameter, etc.).
- **C5.** On-chain governance: Governor / GovernorBravo / OZ Governor / Aragon Voting with token-weighted voting? Record proposal threshold, voting period, quorum, and the timelock delay between queue and execute. Every numeric constant must come from a Read Contract call you can link to, or be in unknowns[] with the C-code. If votingDelay / votingPeriod are denominated in BLOCKS, convert to seconds at the chain's CURRENT block time (Ethereum mainnet ≈ 12s post-Merge, not the 15s in older Compound/Bravo deployments) — cite both block count and converted seconds.
- **C6.** EMERGENCY POWERS: separate emergency-pause / guardian role with a different time cap or different actor than the main upgrade authority? Record it explicitly.
- **C7.** POWER TIER (blast radius). For each privileged path in C3–C6, classify the WORST thing that path can do, choosing the highest applicable tier. Cite the specific function name and any on-chain bound — tier claims without a named function are unsupported.
    - **T1 — FUND-CRITICAL**: replace implementation of contracts holding user funds; change AMM math / accounting / collateral logic; mint unbacked debt or shares; pause withdrawals; drain user-fund treasury; change oracle to attacker-controlled source; replace upgrade admin with EOA.
    - **T2 — ECONOMICALLY MATERIAL**: change fee parameters within bounded ranges; redirect protocol fees; add/remove markets / collateral types; bounded inflation or token mint within hard-capped schedule; spend protocol-owned (non-user) treasury.
    - **T3 — GOVERNANCE-INTERNAL**: change voting rules, quorum, voting period, proposal threshold; upgrade the Governor itself; rotate Timelock admin; mint governance tokens within a capped annual schedule.
    - **T4 — OPERATIONAL**: incentives distribution, grants, ENS / frontend canonicalization, deployment coordination, periphery router deprecation.
  The grade is set by the HIGHEST tier reachable on the uncontested fast path, not the median. State the tier and the binding function in the verdict.

### Read Contract discipline (applies to C3, C4, C5)

Every numeric constant cited (timelock delays, voting periods, multisig thresholds, quorum percentages) must come from EITHER (a) a block-explorer Read Contract URL, OR (b) a DeFiPunkd `/api/contract/read` or `/api/safe/owners` URL (preferred with `&block=<n>` for content-addressed evidence), OR (c) an unknowns[] entry with the C-code. Docs / blog posts are corroboration only — they cannot be the sole citation for a value that is also readable on-chain.

### Off-chain-only substitute hierarchy (when grading_basis="off-chain-only" — see preamble Rule 16)

When on-chain reads were genuinely unreachable this run, eligible off-chain substitutes in priority order:
1. Linked audit PDFs (admin roles, multisig members, timelock delays usually enumerated).
2. Governance forum posts that quote constants from a successful on-chain proposal (cite post URL + linked execution-tx URL).
3. Official protocol docs pages with named addresses and roles (must be on a domain owned by the protocol).
4. GitHub README / SECURITY.md / governance/*.md at a pinned commit SHA.

Forbidden substitutes: third-party blog posts, X / Twitter threads, search-result snippets, model memory. Required degradation: any C-code citing a numeric constant from docs/forum/audit prose ONLY must also carry an `unknowns[]` entry with `-offchain` suffix noting "value not re-read on-chain in this run; corroboration only".

### Grade rules (apply the timelock bar conditional on the highest C7 tier reachable on the fast path)

Security Council standard (used below): a multisig qualifies as "Security Council" only if ALL of: ≥7 signers, ≥51% threshold, ≥50% non-insider signers, every signer publicly announced. Failing any criterion = NOT a Security Council, regardless of signer reputation.

- **green**: highest reachable tier is T3 or T4 regardless of timelock; OR T2 reachable with uncontested-fast-path delay ≥7 days; OR T1 reachable only via immutable contracts (T1 is unreachable); OR T1 reachable with uncontested-fast-path delay ≥7 days combined with a Security Council multisig; OR T1 reachable with uncontested-fast-path delay ≥7 days through active on-chain governance with broad token distribution.
- **orange**: T2 reachable with uncontested-fast-path delay >0 but <7 days; OR T1 reachable with uncontested-fast-path delay >0 but <7 days; OR a multisig failing one or more Security Council criteria sits on a T1/T2 path; OR unclear upgrade authority on a T1/T2 path; OR governance with very short timelock or low quorum on a T1/T2 path.
- **red**: T1 reachable with no timelock by a single EOA or 2-of-3 multisig; OR a T1 upgrade admin that is not a smart contract you can audit.
- **unknown**: completed the checklist but still cannot determine the upgrade authority OR cannot classify the highest tier reachable on the main contracts.

Tiering caveats:
- "Bounded" must be enforced ON-CHAIN to count as T2. A function that sets fees with no upper-bound check is T1 — cite the bound check.
- Recurring T2 economic extraction (e.g. fee redirect with no rate limit) approaches T1 over time. A single proposal that can permanently redirect all future revenue is T1.
- T3 assumes the governance contract cannot itself authorize a T1/T2 action without going through the same timelock. If governance can self-upgrade to bypass the timelock, T3 collapses into T1.
- Do not downgrade tier by hand-waving ("realistically governance would never…"). Tier on what the contract permits, not what feels likely.

Notes:
- **Dynamic / dual-governance timelocks** (Lido, Compound escrow veto): the rubric grades on the uncontested path because that is the path most upgrades take. A dynamic extension that fires only under stake-weighted opposition is a real protection — name it in the green steel-man, but it does not lift an orange fast path into green; state the tension in the verdict.
- **Immutable cores with upgradable governance** (Uniswap-style): if fund-holding contracts are immutable and have no admin-reachable function moving / freezing / re-routing user funds, the highest reachable tier on the upgrade path is T3 — green regardless of timelock. Don't grade this orange just because the Governor is a proxy — that's a C2 fact, not a downgrade. Downgrade only applies if you can cite a concrete function on the upgradable surface that reaches T1 or T2 (privileged hook, upgradable factory controlling fund-routing, fee-switch redirecting protocol revenue without bound).
- **The 7-day bar** reflects the exit-window standard — users need notice after a queued upgrade to withdraw if they disagree. The ability-to-exit slice grades the exit side; this slice grades the delay side; both must hold for users to actually benefit from the delay.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "control",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Ability to exit unknown Unverified

Not yet assessed

Whether users can exit on their own terms if the team disappears or acts adversarially. Requires per-protocol review; not available at Phase 0.

No model has graded this dimension yet. Run the slice prompt through any LLM and submit the JSON — once ≥3 independent runs agree, the quorum bot merges the verdict here.

Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v2
- protocol.name:              Balancer V2
- protocol.chains:            Ethereum, Polygon, Arbitrum, Base, xDai, Avalanche, Polygon zkEVM, Mode, Fraxtal
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/balancer/balancer-v2-monorepo/tree/master/audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: ABILITY-TO-EXIT

Evaluate whether users can withdraw their funds on their own terms, even under adversarial admin conditions.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- E1. Enumerate every user-facing exit function in the main contracts: withdraw, redeem, burn, requestWithdrawal, claim, exit, etc. List them by name. Do NOT treat the contract as a monolith.
- E2. For EACH exit function in E1: identify its access modifiers and any pause guards (e.g. _checkResumed, whenNotPaused, onlyRole). Functions that gate REQUEST PLACEMENT often differ from functions that CLAIM FINALIZED FUNDS — check both separately.
- E3. For each pause guard: identify the role holder (which address holds PAUSE_ROLE / GUARDIAN / etc.) and the maximum pause duration. Specifically check whether PAUSE_INFINITELY (or equivalent uncapped pause) is callable, and which actor can call it (single multisig vs governance vote). For role-holder reads use https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=hasRole&args=0x...,0x... or &method=getRoleAdmin&args=0x.... For "is currently paused" checks use &method=paused or &method=isPaused&args=<resume-code>. Use the BARE method name (no parens). Cite the URL with &block=<n> in evidence[].
- E4. EMERGENCY vs GOVERNANCE pause distinction: many protocols have a fast-acting emergency pause capped at N days and a slow governance pause that can be indefinite. Record both paths separately if present, with their time caps and actor classes.
- E5. Queued redemption: documented maximum queue duration, daily withdrawal caps, whether the queue itself is pausable.
- E6. Forced-exit / escape-hatch / permissionless emergency-exit mechanism for adversarial-admin scenarios.
- E7. Frontend dependency: confirm exit functions are directly callable on-chain (e.g. via Etherscan write tab or a generic wallet) without the project's frontend.

Then write the steel-man section per Hard Rule 11. Common red-vs-orange tension on this slice: indefinite pause exists (suggests red) BUT the realistic emergency path is time-capped AND claims of already-finalized exits are not pause-gated (suggests orange). Resolve this by stating who can do what for how long, not by stopping at the worst-case sentence.

Grade rules:
- green   = permissionless exit; pause is either absent, narrowly scoped to clearly-described emergencies with auto-expiry, or capped at ≤7 days; no frontend dependency for exit; claims of already-finalized exits are not pause-gated under any path.
- orange  = pausable with broad scope OR indefinite pause is reachable only through governance vote (not unilateral admin action), OR queued redemption with documented max > 7 days, OR claims-of-finalized are exempt but new-request placement can be paused indefinitely by governance.
- red     = exit requires admin signature, OR ANY actor (including governance) can pause CLAIMS of finalized exits indefinitely, OR there is no on-chain exit function at all (purely custodial), OR pause is held by a single EOA / 2-of-3 multisig with no time cap.
- unknown = checklist incomplete after checking the sources above.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "ability-to-exit",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Autonomy unknown Unverified

No Phase-0 autonomy signal

Neither the category heuristic nor the forkedFrom signal fires for this protocol. A real autonomy graph (oracles, bridges, fallbacks, governance-mutable dependencies) arrives with Phase-2 onchain discovery.

No model has graded this dimension yet. Run the slice prompt through any LLM and submit the JSON — once ≥3 independent runs agree, the quorum bot merges the verdict here.

Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v2
- protocol.name:              Balancer V2
- protocol.chains:            Ethereum, Polygon, Arbitrum, Base, xDai, Avalanche, Polygon zkEVM, Mode, Fraxtal
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/balancer/balancer-v2-monorepo/tree/master/audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: AUTONOMY

Evaluate this protocol's autonomy: can a failure of anything outside its own contracts cause theft or loss of user principal, loss of unclaimed yield, or materially change the protocol's expected performance? "Autonomy" is not "has zero external touchpoints" — it is "external failures are either survivable or recoverable without user loss". This slice adapts DefiScan v1's Autonomy dimension; dependencies are one of several criteria, not the whole frame.

GUARDRAILS (read before grading):
- Category alone (Liquid Staking, Bridge, RWA Lending, Restaking, …) does NOT force a grade. A category is a hint about where to look; the grade must come from the concrete A1–A9 findings below.
- Base-chain consensus (Ethereum PoS, the chain's validator set, the canonical Deposit Contract at 0x00000000219ab540356cBB839Cbe05303d7705Fa) is the SUBSTRATE, not a dependency, for any protocol deployed on that chain. Do not list "depends on Ethereum" as a finding.
- Oracles or other integrations used by DOWNSTREAM protocols that happen to read this protocol's token (e.g. Chainlink stETH/USD consumed by Aave) are NOT this protocol's dependencies. Count only what THIS protocol's contracts call or trust on-chain. If a feed or contract appears in the protocol's docs only as reference material for third-party integrators, EXCLUDE IT ENTIRELY — do not log it as a finding even with a "peripheral" or "referenced only" caveat.
- "Upgradeable admin can change things" belongs to the CONTROL slice; only count it here when the upgrade surface lets governance silently swap an external dependency (see A9).
- Underlying-asset risk in opt-in, isolated markets is NOT autonomy-red on its own. When a protocol wraps third-party yield-bearing assets (LSTs, LRTs, ERC-4626 vaults, lending receipts, restaked tokens) into per-market silos that users explicitly choose, a failure of one underlying does NOT propagate to other markets and is risk the user opted into per-market. Record it under A4 with depth and propagation scope, but do not let it alone drive a red verdict — red requires an external dependency that cross-cuts the protocol or that the user did not opt into at deposit time. A failure mode that is "if the LRT you deposited is hacked, your principal in that LRT-backed market is impaired" is the underlying's autonomy story, not this protocol's; grade it on whether THIS protocol introduces additional dependencies on top.
- Sub-module enumeration is mandatory before grading. If the protocol ships distinct product lines or modules (e.g., a v2 core AMM plus a newer perps/funding-rate module, a lending pool plus a separate vault layer, an L1 core plus a cross-chain extension), enumerate each in findings and grade against the WORST module weighted by its share of TVS or its blast radius. A green core does not rescue an orange/red sub-module; conversely, a small red sub-module with capped TVS may bound the overall grade to orange. Name each module by its on-chain factory or router address. If you do not know whether a module exists, that is an unknowns[] entry, not silence. EXPLICIT TVS WEIGHTING: in the verdict, state each module's approximate share of total protocol TVS (use "~X%" if exact figures unavailable; check DeFiLlama or block-explorer balances on the module's main contract) and how that share informs the weighted grade. Format: "Module A holds ~X% of TVS (grade: <g>); Module B holds ~Y% (grade: <g>); weighted overall = <grade> because <reason>." If a red sub-module holds <5% of total TVS and is capped, the overall grade may be orange; if it holds >25%, the overall grade is red. Do not let qualitative reasoning substitute for the percentages — write the numbers.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. External contract calls. Enumerate every external contract the core contracts call or read from (oracles, price feeds, AMM pools, lending pools, staking/deposit contracts, yield wrappers). For each, identify the address, the provider, and what user-facing function of this protocol would break or mis-price if that external contract paused, mis-reported, or behaved adversarially. Grep for "oracle", "aggregator", "getPrice", "latestAnswer", "chainlink", "pyth", "redstone" as a starting point. To verify an oracle is live and what it currently reports, hit https://defipunkd.com/api/contract/read?chainId=<id>&address=<oracle>&method=latestAnswer (or &method=latestRoundData, &method=getPrice, &method=description; use the BARE method name without parens) and cite the URL — the response includes blockNumber/blockHash and rawReturnData, which is stronger evidence than a docs page about which feed "should" be used.
- A2. Off-chain actor committees reporting INTO the protocol. Oracle committees, guardian multisigs, DAO-selected validator sets acting as protocol reporters, exit-bus signers, fraud-proof challengers. For each, record committee size, quorum, who picks the members, and what mis-reporting could do (mint, burn, finalize withdrawal, freeze). Distinguish this from governance admins (control slice). NOTE ON STAKING PROTOCOLS: validator slashing and node-operator misbehavior are properties of the base-chain substrate, NOT external dependencies, when the operator set is diversified enough that a coordinated failure caps at <5% principal loss. Count validator/operator risk under A2 ONLY if the operator set is small, non-diversified, or lacks bonding / slashing-insurance / diversification mitigations. A curated set of 30+ independent operators with documented diversification falls under the mitigated path; a 3-operator LST does not. Do not cite the protocol's own risk disclosure as evidence that operator failure = principal loss unless you also check the diversification and bond mitigations.
- A3. Bridge / cross-chain messaging dependencies. Only count bridges that carry material TVL or are required for a core user flow. For each, name the bridge operator (canonical L1↔L2, LayerZero, Wormhole, Axelar, custom multisig), the trust model (canonical, optimistic, light-client, guardian set), and what fraction of TVL or users ride on it. "wstETH exists on 15 chains" is not a finding unless material TVL sits there. Before listing any non-primary chain deployment as a dependency, verify it is still operational as of analysis_date — retired or sunset deployments (e.g., Lido-on-Terra, Lido-on-Solana) belong in unknowns[] or should be omitted, never cited as a current dependency.
- A4. Nested collateral / restaking chains. For restaking / LRT / receipt-of-receipt designs: record the depth of the collateral chain, every actor with slashing or freezing power at each level, and whether a failure N levels deep propagates to user principal here.
- A5. Fork lineage (silent check). If DeFiLlama's forkedFrom is non-empty, record it as one finding and move on. If empty, do NOT add a placeholder finding; it adds noise.
- A6. Fallback mechanisms and circuit breakers. What catches an external failure? Sanity-check contracts on oracle reports, rebase bounds, pause paths triggered by bad prices, second-opinion oracles, max-per-block throughput caps, withdrawal queues that absorb bad reports. Record which A1–A4 risks are mitigated by which fallback, and which are unmitigated. For EACH fallback you cite, state its activation status explicitly: (i) LIVE and enforcing on-chain today, (ii) DEPLOYED but not yet wired / activated (e.g., interface exists but the address is zero or the role is unassigned), or (iii) DOCUMENTED / PROPOSED only (forum post, LIP draft, audit pending). Only (i) counts as mitigation for the grade. A fallback in state (ii) or (iii) should be noted but must not reduce the risk in your steel-man or verdict. If you cannot determine activation status, add an unknowns[] entry rather than assume it is live.
- A7. Sequencer / L1-liveness dependency BEYOND the base-chain substrate. SCOPE — sequencer risk only counts here when the protocol IS its own L2/L3 appchain or app-rollup, where the sequencer is part of the protocol's own stack and a freeze is a protocol-level outage. A protocol permissionlessly deployed on a third-party L2 inherits that L2's sequencer as substrate, not an A7 dependency. Record the sequencer/DA trust model when A7 applies.
- A8. Keeper / relayer / off-chain bot liveness. Protocols that need permissionless-but-necessary off-chain actors (liquidation bots, auto-compounders, deposit relayers, intent solvers). Record whether the role is permissionless, what degrades if nobody runs it (yield paused, bad debt accumulates, positions go stale), and whether the failure mode is graceful or catastrophic.
- A9. Governance-mutable dependency surface. Can an admin or DAO action silently INTRODUCE a NEW EXTERNAL dependency — swap the oracle address to a different provider, register a new staking module that calls an untrusted contract, add a new bridge, route SY through a new external vault — without an exit window for users? Check the upgrade / router / module-registry contracts. Answer: which EXTERNAL dependencies are governance-mutable, who holds that power, and whether there is a timelock or exit window. SCOPE LIMIT — read carefully: A9 is about the *external dependency surface*, not the upgrade surface in general. "The proxyAdmin / EOA can upgrade the router implementation to arbitrary bytecode" is a CONTROL-slice finding (admin can rug), NOT an autonomy-A9 finding. A9 fires only when the upgrade specifically swaps out or adds a contract that THIS protocol calls or trusts (e.g., changing the oracle address from Chainlink to a malicious feed, registering a new SY adapter that points to a third-party vault, redirecting a bridge endpoint). If the only finding is "admin can change implementation," do not log it under A9 and do not let it drive the autonomy grade — note it under control instead. The autonomy-relevant version of the same upgrade key is "admin can swap [specific external address X] without timelock"; that requires identifying the specific external dependency that becomes mutable.

STEEL-MAN (per Hard Rule 13): write one-sentence strongest arguments for red, orange, and green using the A1–A9 findings.

IMPACTED TVS ESTIMATE: the headline MUST include a rough impacted-TVS figure — the fraction of protocol TVS that could be lost or frozen if the worst-unmitigated dependency you identified failed. Use "~X%" if exact numbers are unavailable, "<1%" for de minimis, "unclear" only if A1–A9 left the question genuinely open (in which case grade=unknown is usually correct). Do NOT substitute qualitative phrases like "significant" — give a number or bracket.

GRADE ANCHORS (mapped to DefiScan v1 stages):
- green = Stage 2 equivalent. Failure of any external dependency cannot cause loss of user principal or unclaimed yield. Either there are no material external dependencies, or every critical one has a documented fallback (A6) that keeps users whole. Governance cannot silently introduce new dependencies without an exit window (A9). Impacted TVS under any single-dependency failure: effectively 0.
- orange = Stage 1 equivalent. Failure of some external dependency can cause loss of unclaimed yield, or can materially change expected performance (pause withdrawals, freeze positions, degrade price quality), but cannot cause loss of principal. Committee-based oracles with sanity checks, canonical-only bridges, fallback paths that exist but are incomplete, or governance-mutable dependencies protected by a ≥7-day timelock. Impacted TVS is bounded and recoverable.
- red = Stage 0 equivalent. Failure of an external dependency CAN cause theft or loss of principal. Examples: single-provider oracle with no sanity check or fallback, material TVL on a non-canonical bridge with a guardian multisig, governance can hot-swap oracles or add staking modules with no timelock or exit window, unmitigated keeper-liveness dependency where positions become insolvent if bots stop. Impacted TVS is material.
- unknown = checklist incomplete after inspecting source + verified contracts. Prefer unknown over guessing when A1/A6 could not be reconstructed. RESERVE unknown for cases where the CORE ARCHITECTURE itself is unverifiable — not for cases where you merely cannot enumerate every per-market dependency in a multi-market protocol. If the core router/factory/oracle architecture is verifiable on-chain and you can determine whether the core requires external dependencies, grade the architecture even when an exhaustive per-market external-dependency census is infeasible. Acknowledge the per-market gap in unknowns[] but still issue a grade. Refusing to grade a multi-market protocol because you cannot list every SY/vault/market is over-use of unknown; grade the architecture and say so.

PROMPT-META CHECK (per Hard Rule 17): before finalizing, verify the verdict cites concrete contract addresses, docs, or code — not the rubric itself. If your verdict says "the protocol belongs to a category the rubric marks red", rewrite it with the A1–A9 finding that actually justifies the grade, or drop to grade=unknown.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "autonomy",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Open Access unknown Unverified

Not yet assessed

Whether the protocol depends on privileged operators, whitelists, geo-restrictions, or off-chain infrastructure. This is not a signal DeFiLlama carries in a usable form; crawler-based detection lands in a later phase.

No model has graded this dimension yet. Run the slice prompt through any LLM and submit the JSON — once ≥3 independent runs agree, the quorum bot merges the verdict here.

Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v2
- protocol.name:              Balancer V2
- protocol.chains:            Ethereum, Polygon, Arbitrum, Base, xDai, Avalanche, Polygon zkEVM, Mode, Fraxtal
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/balancer/balancer-v2-monorepo/tree/master/audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: OPEN-ACCESS

Evaluate who is allowed to use the protocol and whether any of that permission is granted off-chain.

Scope: this slice is about ADMISSION — who can enter, exit, or transact. Operator LIVENESS (what breaks if keepers/oracles go offline) is assessed in the dependencies slice and is out of scope for the grade here. You may note operator dependencies as context, but do not let "the protocol halts if operator X disappears" drive the access grade on its own; that belongs in dependencies. Source verification / contract verification on block explorers is assessed in the verifiability slice and is out of scope here — do NOT let "contract is unverified" drive the access grade.

Framing: the smart contracts are the access layer; frontends are UX. A permissionless contract is reachable by any client (SDK, third-party UI, aggregator, wallet integration). Frontend ToS, IP geo-blocking, and wallet screening are publisher policies on one specific client — they are reported as context but do NOT determine the grade. The grade hinges on (1) what the contract itself permits, and (2) whether the protocol is practically reachable without the official publisher's cooperation.

Meta-check before finalizing: if your verdict cites phrases from this prompt as evidence ("the protocol meets the 'credible alternatives' condition", "this fits the 'documented fallback' rule"), redo the verdict. The prompt describes the rubric; evidence must come from the protocol. A verdict should cite what the protocol does, not what the rubric says.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. Whitelist / allowlist modifiers in user-facing entry points. Grep for "onlyWhitelisted", "onlyRole", "allowlist", "isAccredited", "isKYCed". Note which functions are gated and who can add/remove from the list.
- A2. Off-chain operators in the admission path: keepers, sequencers, privileged relayers, oracle posters whose approval is required to admit a user action (not just to keep the protocol live). For each, identify whether the role is held by a single operator, a permissioned committee, or is permissionless. Enumerate per user-facing function class (deposit vs withdraw-request vs claim-finalized vs transfer) which ones require operator approval to be admitted, and which ones admit users unconditionally. A function whose placement is unconditional but whose downstream settlement depends on an operator is an admission-permissionless function — flag the liveness dependency as context and defer its grading weight to the dependencies slice.
- A3. Frontend restrictions on the official interface — record as context, not as a grade lever. Distinguish:
    - A3-passive: boilerplate ToS clauses (sanctions attestation, restricted-territory self-certification, VPN-circumvention prohibition, "comply with applicable law" eligibility, age of majority).
    - A3-active: runtime enforcement — IP-based geo-blocking, wallet-address screening against a sanctions oracle (Chainalysis, TRM, Elliptic), KYC wall, rendering-blocking jurisdiction banner.
  Record findings under the correct tier. Quote ToS text or banner text in evidence[].shows. These findings populate the headline and rationale but do NOT move the grade by themselves; the grade is set by A1, A2, and the A3b path check below.
- A3b. Independent access paths (the operative grade input). Enumerate paths that do not require the official publisher's cooperation:
    - Published SDK / library / CLI for direct contract interaction.
    - Third-party frontends operated by separate legal entities.
    - Wallet-integrated access (MetaMask Swaps, Safe apps, etc.).
    - DEX / lending / yield aggregators that route through the contracts.
  Record at least one concrete link per path that exists. The protocol does NOT have to self-document these — the test is existence, not UX cost. An A3b-i redistribution of the official UI bound by the same ToS does NOT count as an independent path.
- A4. Sanctions / compliance tooling at the contract level: does the protocol check addresses against OFAC lists or similar on-chain blocklists in the contract itself? (Frontend-only screening belongs in A3.)
- A5. Differentiate read access vs write access: many protocols are read-permissionless (anyone can view state) but write-gated (only certain addresses can deposit/borrow). Record both.
- A6. ToS / Legal links: locate them on the website and produce a VERBATIM quote of any jurisdictional, sanctions, or eligibility clause in evidence[].shows. If you cannot extract the clause text verbatim (SPA render failure, paywall, dead link, etc.), do NOT paraphrase or infer from general knowledge — record the ToS URL in unknowns[] with the reason extraction failed. Assertions about ToS content without a verbatim quote will be downweighted by reviewers.

Then write the steel-man section per Hard Rule 11.

Grade rules (admission-focused; liveness concerns belong in dependencies; source verification belongs in verifiability):
- green   = no contract-level whitelist/KYC on user entry/exit; no operator approval required to admit a user action; AND at least one independent A3b path exists (published SDK, third-party frontend, wallet integration, or aggregator routing). Frontend ToS posture and A3-active enforcement on the official UI do NOT block green when contracts are permissionless and an independent path exists — they are reported as context.
- orange  = contracts admit users unconditionally, BUT the protocol is operationally captured by the official publisher: no published SDK, no third-party frontend, no wallet integration, no aggregator routing. The contract is theoretically open but practically reachable only through the official UI. Also applies when admission requires approval from a permissioned committee that is governance-managed with a documented replacement procedure.
- red     = contract-level whitelist / KYC on user entry/exit, OR admission of a core user action requires approval from a single privileged operator or a small committee with no documented replacement procedure, OR enforces an on-chain blocklist updatable by a single party.
- unknown = checklist incomplete after checking the sources above.

Default-grade guidance: when contracts are fully permissionless AND any A3b independent path exists, the default grade is green regardless of frontend ToS or A3-active enforcement on the official UI. Frontend geo-blocking, sanctions-oracle wallet screening, and ToS sanctions clauses are publisher policies on one client and are reported in findings/headline as context, not as grade levers. To grade orange on operational-capture grounds, the auditor must affirmatively show that ALL independent paths are absent or also gated.

Guideline on committees: where admission depends on a multi-operator committee, the relevant axes are (a) set size, (b) whether replacement/rotation is governed on-chain, (c) whether the replacement procedure is publicly documented. A large set with on-chain governance replacement should not be graded as a single-party operator even if rotation is not instantaneous. A small set with informal replacement should be treated as a single-party operator.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "open-access",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Stage

Preview of the Phase-3 maturity framework. DeFiPunk'd will adopt DeFiScan v2's stages verbatim; the section is rendered below in its intended shape so the structure is visible today.

Balancer V2 has not yet been assessed under the DeFiScan v2 stage framework.

The walkaway test is the central criterion. Once stages land, protocols reach Stage 1 only if users can exit in the presence of malicious operators even when the emergency council disappears.

Scope of assessment

Stage 0 requirements pending

Stage 1 requirements pending

Stage 2 requirements pending

Learn more about DeFiScan v2 stages →

TVL $7.2M

Type Dexs

Chain Ethereum

View on DeFiLlama ↗

Control Exit Autonomy Open Access Verifiable

Verifiability tentative Open source + 3 audits
Control unknown Not yet assessed
Ability to exit unknown Not yet assessed
Autonomy unknown No Phase-0 autonomy signal
Open Access unknown Not yet assessed

Control criteria

Upgradeability — Bug bounty immunefi.com Governance forum — Docs —

About

Balancer V1 is a dexs protocol on Ethereum.

Risk analysis

One card per dimension, sorted by severity. Only Verifiability and Autonomy carry automated signals in Phase 0. See methodology for scope.

Audit a dimension yourself · DEFI@home Contribute an LLM-run assessment — any model, any dimension. Three agreeing runs merge automatically into the public record.

Address discovery 0 addresses on file · 0 runs ⚑ Run first Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v1
- protocol.name:              Balancer V1
- protocol.chains:            Ethereum
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://docs.balancer.fi/reference/contracts/security.html#audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: DISCOVERY

You are a cataloguer, not a judge. Your job is to surface every contract address that could plausibly belong to this protocol's control or fund-holding surface, each backed by a citation. `grade` is ALWAYS `"unknown"` for discovery submissions — there is no green/orange/red rubric here. The five evaluation slices that run after you (control, ability-to-exit, autonomy, open-access, verifiability) consume your output via the addressBook ratchet — every address you record becomes a pre-built surfacer URL on the next run; every address you miss costs them a tool call.

Width beats depth. A `role: "other"` entry with one cited URL beats omitting it. Downstream slices will discard out-of-scope entries; they cannot rediscover what you fail to enumerate without paying the same cost again.

(Step 0 capability probe lives in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every D-code below must appear in evidence[] OR unknowns[])

- **D1. Block-explorer name-tag search per chain.** For each chain in `protocol.chains`, search the canonical block explorer for the protocol's name tag — `https://etherscan.io/searchHandler?term=<query>` and the per-chain explorers (basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network). When direct fetch is blocked, use `site:<explorer> <protocol_name>` via search grounding. Record every address that surfaces with the protocol's tag, plus neighbouring "Token Contract" / "Multisig" / "Timelock" labels.

- **D2. Official deployments doc.** From `protocol.website` and the docs site, locate the canonical "Deployed contracts" / "Addresses" / "Contracts" / "Deployments" page (often at `/docs/deployments`, `/docs/addresses`, `/dashboard/contracts`). Cite the URL, record every address listed with its named role, and set `protocol_metadata.deployed_contracts_doc` to this URL.

- **D3. Audit PDFs.** From `protocol.audit_links` (and any audits surfaced by D2), open each. Most reports include a "Scope" / "Contracts in scope" address table in the first 5 pages. Extract every in-scope address with its labelled role. If the audit predates the current deployment, record the addresses anyway with role suffixed `(audit-era)` so downstream slices know to re-verify.

- **D4. GitHub deployment artifacts.** From `protocol.github`, walk the repo at a pinned commit SHA looking for: Foundry `broadcast/<script>/<chainId>/run-latest.json` (`transactions[].contractAddress` per chainId); hardhat-deploy `deployments/<network>/<Contract>.json` (`address` field); manual indexes (`deployments.json`, `addresses.json`, `contracts.json`, `networks.json`); markdown indexes (`docs/deployments.md`, `README.md` tables). Cite the file URL with the commit SHA; pin SHAs (`?ref=<sha>`) so the citation is content-addressed.

- **D5. Multi-chain enumeration.** If `protocol.chains.length > 1`, repeat D1–D4 per chain. Cross-chain deployments of the same logical contract get SEPARATE `admin_addresses[]` entries — one per chain. The chain field is part of the identity; do not collapse. If a chain has zero results, record `"D5: chain <name>: zero addresses surfaced from <sources tried>"` in unknowns[].

- **D6. Factory-discovered children.** For factory addresses surfaced in D1–D4, fetch the enumeration view via the read API (`/api/contract/read?...&method=allPools` / `getPool` / `getMarket` / `getVault`) and record each child with role like `"pool (from factory <0xFactory>)"`. **Cap at 50 children per factory.** Protocols with thousands of pools (Uniswap, Sushi) need dedicated ingestion — record the factory + the cap notice in unknowns[].

- **D7. Role taxonomy.** Every `admin_addresses[]` entry's `role` uses this controlled vocabulary (free-text suffixes OK for disambiguation, e.g. `"multisig (treasury)"`, but the leading token must match):

    `owner | admin | proxy_admin | governor | timelock | guardian | multisig | treasury | oracle | factory | router | token | pool | vault | other`

  Tentative classifications are encouraged. `actor_class` ∈ `eoa | multisig | timelock | governance | unknown` — use `unknown` when you found the address but didn't read its bytecode.

- **D8. Ratchet output integrity.** Every address in `admin_addresses[]` must trace to ≥1 fetched URL in evidence[]. Snippet-only sightings go in unknowns[] with a `D8` code, NOT in admin_addresses[].

### Discovery rationale framing

- `rationale.findings`: one entry per D-code, terse, factual. Per-address detail belongs in evidence[] and admin_addresses[], not here. Example: `"D1: 8 addresses surfaced from etherscan.io name-tag search for 'Aave V3'"`.
- `rationale.steelman`: ALWAYS null.
- `rationale.verdict`: one short line summarizing what corpora were walked and how many addresses were catalogued.
- `headline`: factual and quantitative — `"24 contracts catalogued across Ethereum, Arbitrum, and Base; 6 governance/admin and 18 protocol contracts."`.
- `short_headline`: under 60 chars — `"24 contracts across 3 chains"`.

### What discovery is NOT

- Not a verdict slice. `grade` must be `"unknown"`.
- Not exhaustive enumeration of leaf assets — record the factory + cap and move on (see D6).
- Not classification of trust assumptions — whether a multisig threshold is safe / timelock delay is sufficient / proxy admin is an EOA is the control slice's job.
- Not address-book reconciliation: when addressBook is non-empty, EXTEND it (find addresses prior runs missed) rather than re-cite the same addresses; re-cite only when you have new evidence for a refined role.

### protocol_metadata side-effects

While walking the corpora, populate every `protocol_metadata` field you can support with citations: `github`, `docs_url`, `audits` (one per D3 audit walked), `governance_forum`, `bug_bounty_url`, `security_contact`, `deployed_contracts_doc` (URL from D2), `upgradeability` (best-effort), and `about` (2–4 sentences sourced from docs/website, not memory). Discovery is the natural home for these — evaluation slices should not have to rediscover them.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "discovery",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Verifiability Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v1
- protocol.name:              Balancer V1
- protocol.chains:            Ethereum
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://docs.balancer.fi/reference/contracts/security.html#audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: VERIFIABILITY

Evaluate whether an outsider can independently confirm what the deployed code does.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- V1. For each address you assess: is the bytecode verified on the chain's block explorer? Record the "Contract Source Code Verified" indicator. https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x... returns this as a top-level "verified" boolean plus "abiSource" ("etherscan" / "sourcify") and an inline ABI — useful when the explorer page is rate-limited. If the contract is a proxy, verify BOTH the proxy contract AND the current implementation contract. The same /api/contract/abi response auto-resolves proxies and includes a "proxy.implementation" address when present, so one call covers V1 + V6 in one shot. An explorer "Similar Match" on a well-known proxy pattern (Aragon AppProxyUpgradeable, ERC1967Proxy, OssifiableProxy, OZ TransparentUpgradeableProxy) is expected for that pattern and does NOT count as a verification gap on its own — what matters is that the implementation is independently verified.
- V2. Source-to-repo correspondence: for each verified contract, attempt to find a matching commit in the linked GitHub repos. Record evidence[].commit on a match. Independent compile/bytecode-match is NOT required for green — a recognized public repo whose structure and file contents correspond to the explorer-visible source is sufficient. If you did not pin a commit SHA or run a bytecode diff, record that plainly in unknowns[] and proceed; it is a scope limit, not a downgrade signal.
- V3. Audit coverage: for each URL in protocol.audit_links, open it and record: auditor name, audit date, the specific contracts / commit in scope. Flag audits that predate the current deployment by >6 months without a follow-up review.
- V4. Auditor recognition: the following firms are broadly recognized in Solidity: Trail of Bits, Zellic, Spearbit, OpenZeppelin, ConsenSys Diligence, Certora (formal verification), Quantstamp, Halborn, Peckshield, Sigma Prime, ChainSecurity, Ackee Blockchain, MixBytes, Statemind. Unknown firms are orange-at-best for any green-grade claim. Name the firm explicitly in evidence[].
- V5. Post-audit drift: compare the most recent in-scope audit(s) against the currently-deployed source, weighted by what each contract does and by what the changes actually contain. SCOPE — drift only downgrades the grade when ALL of the following hold: (i) the drifting contract is fund-custody / settlement / accounting-critical (NOT a peripheral router, lens, quoter, or pure-view contract that holds no balances); (ii) the changes are material — new functions, modified access control, modified accounting, modified fund flow — and not refactors / struct relocations / import reorgs / build or CI fixes / formatting; (iii) no later audit, fix-audit, or differential audit from a recognized firm covers the changed files (audits often pin a pre-fix commit while a follow-up reviews the delta — match by file scope, not by commit-hash equality). When you cite drift as a downgrade reason, name the specific behavior change (function added, role granted, accounting formula altered) — "N commits ahead" or "+X/-Y LOC" alone is not evidence of material drift. If you have not sampled the diff content (e.g. via the GitHub compare view or the top commits in the window), record drift as an unknowns[] entry rather than auto-downgrading; commit-count and LOC are starting signals, not findings.
- V6. Implementation vs proxy: a verified proxy with an unverified implementation is effectively unverified. State whether the implementation is verified separately.

EVIDENCE DISCIPLINE (read before writing findings[]):
- Do not assert a specific deploy-commit SHA, bytecode equivalence, or "identical to audited commit" unless you actually fetched the artifact that shows it (e.g., a deployed-addresses JSON you opened, an explorer page you read). Inferred or plausible matches belong in unknowns[], never in findings[] or evidence[].
- Evidence[] entries must correspond to pages/files you actually retrieved this run. A URL you did not open is not evidence.

Then write the steel-man section per Hard Rule 11.

Grade rules:
- green   = deployed bytecode verified on the explorer (proxy AND implementation if proxied; "Similar Match" on a standard proxy pattern is fine per V1), a public source repo exists whose contents correspond to the explorer-visible source, AND ≥1 audit from a recognized firm covering the currently-deployed contracts (≤6 months of drift OR drift was re-audited). A missing local compile-match is not a downgrade — record it in unknowns[] and still grade green if the other conditions hold.
- orange  = verified but with visible drift from the public repo, OR audit scope is stale relative to deployment, OR only minor / unknown-firm audits exist, OR only some of the main contracts are verified, OR proxy verified but implementation only partially verified.
- red     = unverified bytecode (or verified proxy with unverified implementation), OR no audit in protocol.audit_links, OR no public repo.
- unknown = reserved for when the protocol's verifiability posture genuinely cannot be assessed (e.g., explorer and repo both inaccessible for this protocol). Do NOT use unknown merely because you, the analyst, could not run a particular check such as a bytecode diff — that goes in unknowns[] while the grade is still assigned from the evidence you do have.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "verifiability",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Control Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v1
- protocol.name:              Balancer V1
- protocol.chains:            Ethereum
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://docs.balancer.fi/reference/contracts/security.html#audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: CONTROL

Evaluate who can change the protocol's rules, how fast, and how broadly.

(Step 0 capability probe and the off-chain-only fallback live in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[])

- **C1.** For each address you assess: who is the contract owner / admin / pendingAdmin / governor — read these via the block explorer's "Read Contract" tab OR `https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=owner` (BARE method names: `&method=owner`, `&method=admin`, `&method=pendingOwner`, `&method=governor`). For Safes use `/api/safe/owners?chainId=<id>&address=0x...`. When a protocol has multiple major versions deployed (v2/v3/v4), perform C1 reads on the NEWEST deployment separately — newer deployments often have weaker control surfaces than the legacy core.
- **C2.** Upgrade mechanism: transparent proxy / UUPS / Beacon / Diamond / immutable. Identify the proxy admin address. Check upgradeability of GOVERNANCE contracts too — a Governor / Aragon Voting / OZ Governor is often itself a proxy whose admin is the Timelock. Asymmetry: when fund-holding cores are immutable AND governance has no admin path that reaches them, an upgradable Governor/Timelock is T3-only and must NOT drag the verdict below green on that basis alone (see grade rules and the "immutable cores" caveat). Only call upgradability "mixed" if you can name a concrete function on the upgradable surface that reaches T1 or T2 on user funds.
- **C3.** EXECUTION PATH (enumerate every stage, in order, with delays in seconds). The operative path is usually a chain (voting → scheduler → timelock → executor; or governor → queue → execute; or Aragon Voting → DualGovernance → EmergencyProtectedTimelock → AdminExecutor). For each stage, record (a) the contract address, (b) the delay constant name + value in seconds, (c) the URL you read it from (block-explorer Read Contract OR `/api/contract/read?...&method=MIN_DELAY&block=<n>`). Do NOT stop at the first timelock-shaped contract — if its admin is itself called by another contract, keep walking. The grading delay is the SUM OF DELAYS ON THE UNCONTESTED FAST PATH (shortest time a proposal with no opposition can go from submission to executable). Dynamic / contested extensions (veto signaling, rage quit, escrow delay) are modifiers, not the basis — note them separately.
- **C4.** Enumerate EVERY multisig with reachable control — main proxy admin, emergency activation, emergency execution, reseal / pause, gate-seal committees, tiebreaker, per-module admins. For each Safe, fetch threshold + owners + version via `/api/safe/owners?chainId=<id>&address=0x...` (response includes raw eth_call data, so the URL is citable evidence). Enumerate ops/council/incentives multisigs even when off the upgrade path — record their scope so a reader can see they are NOT on the upgrade path. For each: (a) address, (b) threshold / total signers, (c) signer identities classified as insider (team, paid auditors under ongoing engagement, mandated service providers) vs non-insider (independent community members, unaffiliated researchers), (d) the specific power held (upgrade, pause, parameter, etc.).
- **C5.** On-chain governance: Governor / GovernorBravo / OZ Governor / Aragon Voting with token-weighted voting? Record proposal threshold, voting period, quorum, and the timelock delay between queue and execute. Every numeric constant must come from a Read Contract call you can link to, or be in unknowns[] with the C-code. If votingDelay / votingPeriod are denominated in BLOCKS, convert to seconds at the chain's CURRENT block time (Ethereum mainnet ≈ 12s post-Merge, not the 15s in older Compound/Bravo deployments) — cite both block count and converted seconds.
- **C6.** EMERGENCY POWERS: separate emergency-pause / guardian role with a different time cap or different actor than the main upgrade authority? Record it explicitly.
- **C7.** POWER TIER (blast radius). For each privileged path in C3–C6, classify the WORST thing that path can do, choosing the highest applicable tier. Cite the specific function name and any on-chain bound — tier claims without a named function are unsupported.
    - **T1 — FUND-CRITICAL**: replace implementation of contracts holding user funds; change AMM math / accounting / collateral logic; mint unbacked debt or shares; pause withdrawals; drain user-fund treasury; change oracle to attacker-controlled source; replace upgrade admin with EOA.
    - **T2 — ECONOMICALLY MATERIAL**: change fee parameters within bounded ranges; redirect protocol fees; add/remove markets / collateral types; bounded inflation or token mint within hard-capped schedule; spend protocol-owned (non-user) treasury.
    - **T3 — GOVERNANCE-INTERNAL**: change voting rules, quorum, voting period, proposal threshold; upgrade the Governor itself; rotate Timelock admin; mint governance tokens within a capped annual schedule.
    - **T4 — OPERATIONAL**: incentives distribution, grants, ENS / frontend canonicalization, deployment coordination, periphery router deprecation.
  The grade is set by the HIGHEST tier reachable on the uncontested fast path, not the median. State the tier and the binding function in the verdict.

### Read Contract discipline (applies to C3, C4, C5)

Every numeric constant cited (timelock delays, voting periods, multisig thresholds, quorum percentages) must come from EITHER (a) a block-explorer Read Contract URL, OR (b) a DeFiPunkd `/api/contract/read` or `/api/safe/owners` URL (preferred with `&block=<n>` for content-addressed evidence), OR (c) an unknowns[] entry with the C-code. Docs / blog posts are corroboration only — they cannot be the sole citation for a value that is also readable on-chain.

### Off-chain-only substitute hierarchy (when grading_basis="off-chain-only" — see preamble Rule 16)

When on-chain reads were genuinely unreachable this run, eligible off-chain substitutes in priority order:
1. Linked audit PDFs (admin roles, multisig members, timelock delays usually enumerated).
2. Governance forum posts that quote constants from a successful on-chain proposal (cite post URL + linked execution-tx URL).
3. Official protocol docs pages with named addresses and roles (must be on a domain owned by the protocol).
4. GitHub README / SECURITY.md / governance/*.md at a pinned commit SHA.

Forbidden substitutes: third-party blog posts, X / Twitter threads, search-result snippets, model memory. Required degradation: any C-code citing a numeric constant from docs/forum/audit prose ONLY must also carry an `unknowns[]` entry with `-offchain` suffix noting "value not re-read on-chain in this run; corroboration only".

### Grade rules (apply the timelock bar conditional on the highest C7 tier reachable on the fast path)

Security Council standard (used below): a multisig qualifies as "Security Council" only if ALL of: ≥7 signers, ≥51% threshold, ≥50% non-insider signers, every signer publicly announced. Failing any criterion = NOT a Security Council, regardless of signer reputation.

- **green**: highest reachable tier is T3 or T4 regardless of timelock; OR T2 reachable with uncontested-fast-path delay ≥7 days; OR T1 reachable only via immutable contracts (T1 is unreachable); OR T1 reachable with uncontested-fast-path delay ≥7 days combined with a Security Council multisig; OR T1 reachable with uncontested-fast-path delay ≥7 days through active on-chain governance with broad token distribution.
- **orange**: T2 reachable with uncontested-fast-path delay >0 but <7 days; OR T1 reachable with uncontested-fast-path delay >0 but <7 days; OR a multisig failing one or more Security Council criteria sits on a T1/T2 path; OR unclear upgrade authority on a T1/T2 path; OR governance with very short timelock or low quorum on a T1/T2 path.
- **red**: T1 reachable with no timelock by a single EOA or 2-of-3 multisig; OR a T1 upgrade admin that is not a smart contract you can audit.
- **unknown**: completed the checklist but still cannot determine the upgrade authority OR cannot classify the highest tier reachable on the main contracts.

Tiering caveats:
- "Bounded" must be enforced ON-CHAIN to count as T2. A function that sets fees with no upper-bound check is T1 — cite the bound check.
- Recurring T2 economic extraction (e.g. fee redirect with no rate limit) approaches T1 over time. A single proposal that can permanently redirect all future revenue is T1.
- T3 assumes the governance contract cannot itself authorize a T1/T2 action without going through the same timelock. If governance can self-upgrade to bypass the timelock, T3 collapses into T1.
- Do not downgrade tier by hand-waving ("realistically governance would never…"). Tier on what the contract permits, not what feels likely.

Notes:
- **Dynamic / dual-governance timelocks** (Lido, Compound escrow veto): the rubric grades on the uncontested path because that is the path most upgrades take. A dynamic extension that fires only under stake-weighted opposition is a real protection — name it in the green steel-man, but it does not lift an orange fast path into green; state the tension in the verdict.
- **Immutable cores with upgradable governance** (Uniswap-style): if fund-holding contracts are immutable and have no admin-reachable function moving / freezing / re-routing user funds, the highest reachable tier on the upgrade path is T3 — green regardless of timelock. Don't grade this orange just because the Governor is a proxy — that's a C2 fact, not a downgrade. Downgrade only applies if you can cite a concrete function on the upgradable surface that reaches T1 or T2 (privileged hook, upgradable factory controlling fund-routing, fee-switch redirecting protocol revenue without bound).
- **The 7-day bar** reflects the exit-window standard — users need notice after a queued upgrade to withdraw if they disagree. The ability-to-exit slice grades the exit side; this slice grades the delay side; both must hold for users to actually benefit from the delay.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "control",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Ability to exit Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v1
- protocol.name:              Balancer V1
- protocol.chains:            Ethereum
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://docs.balancer.fi/reference/contracts/security.html#audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: ABILITY-TO-EXIT

Evaluate whether users can withdraw their funds on their own terms, even under adversarial admin conditions.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- E1. Enumerate every user-facing exit function in the main contracts: withdraw, redeem, burn, requestWithdrawal, claim, exit, etc. List them by name. Do NOT treat the contract as a monolith.
- E2. For EACH exit function in E1: identify its access modifiers and any pause guards (e.g. _checkResumed, whenNotPaused, onlyRole). Functions that gate REQUEST PLACEMENT often differ from functions that CLAIM FINALIZED FUNDS — check both separately.
- E3. For each pause guard: identify the role holder (which address holds PAUSE_ROLE / GUARDIAN / etc.) and the maximum pause duration. Specifically check whether PAUSE_INFINITELY (or equivalent uncapped pause) is callable, and which actor can call it (single multisig vs governance vote). For role-holder reads use https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=hasRole&args=0x...,0x... or &method=getRoleAdmin&args=0x.... For "is currently paused" checks use &method=paused or &method=isPaused&args=<resume-code>. Use the BARE method name (no parens). Cite the URL with &block=<n> in evidence[].
- E4. EMERGENCY vs GOVERNANCE pause distinction: many protocols have a fast-acting emergency pause capped at N days and a slow governance pause that can be indefinite. Record both paths separately if present, with their time caps and actor classes.
- E5. Queued redemption: documented maximum queue duration, daily withdrawal caps, whether the queue itself is pausable.
- E6. Forced-exit / escape-hatch / permissionless emergency-exit mechanism for adversarial-admin scenarios.
- E7. Frontend dependency: confirm exit functions are directly callable on-chain (e.g. via Etherscan write tab or a generic wallet) without the project's frontend.

Then write the steel-man section per Hard Rule 11. Common red-vs-orange tension on this slice: indefinite pause exists (suggests red) BUT the realistic emergency path is time-capped AND claims of already-finalized exits are not pause-gated (suggests orange). Resolve this by stating who can do what for how long, not by stopping at the worst-case sentence.

Grade rules:
- green   = permissionless exit; pause is either absent, narrowly scoped to clearly-described emergencies with auto-expiry, or capped at ≤7 days; no frontend dependency for exit; claims of already-finalized exits are not pause-gated under any path.
- orange  = pausable with broad scope OR indefinite pause is reachable only through governance vote (not unilateral admin action), OR queued redemption with documented max > 7 days, OR claims-of-finalized are exempt but new-request placement can be paused indefinitely by governance.
- red     = exit requires admin signature, OR ANY actor (including governance) can pause CLAIMS of finalized exits indefinitely, OR there is no on-chain exit function at all (purely custodial), OR pause is held by a single EOA / 2-of-3 multisig with no time cap.
- unknown = checklist incomplete after checking the sources above.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "ability-to-exit",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Autonomy Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v1
- protocol.name:              Balancer V1
- protocol.chains:            Ethereum
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://docs.balancer.fi/reference/contracts/security.html#audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: AUTONOMY

Evaluate this protocol's autonomy: can a failure of anything outside its own contracts cause theft or loss of user principal, loss of unclaimed yield, or materially change the protocol's expected performance? "Autonomy" is not "has zero external touchpoints" — it is "external failures are either survivable or recoverable without user loss". This slice adapts DefiScan v1's Autonomy dimension; dependencies are one of several criteria, not the whole frame.

GUARDRAILS (read before grading):
- Category alone (Liquid Staking, Bridge, RWA Lending, Restaking, …) does NOT force a grade. A category is a hint about where to look; the grade must come from the concrete A1–A9 findings below.
- Base-chain consensus (Ethereum PoS, the chain's validator set, the canonical Deposit Contract at 0x00000000219ab540356cBB839Cbe05303d7705Fa) is the SUBSTRATE, not a dependency, for any protocol deployed on that chain. Do not list "depends on Ethereum" as a finding.
- Oracles or other integrations used by DOWNSTREAM protocols that happen to read this protocol's token (e.g. Chainlink stETH/USD consumed by Aave) are NOT this protocol's dependencies. Count only what THIS protocol's contracts call or trust on-chain. If a feed or contract appears in the protocol's docs only as reference material for third-party integrators, EXCLUDE IT ENTIRELY — do not log it as a finding even with a "peripheral" or "referenced only" caveat.
- "Upgradeable admin can change things" belongs to the CONTROL slice; only count it here when the upgrade surface lets governance silently swap an external dependency (see A9).
- Underlying-asset risk in opt-in, isolated markets is NOT autonomy-red on its own. When a protocol wraps third-party yield-bearing assets (LSTs, LRTs, ERC-4626 vaults, lending receipts, restaked tokens) into per-market silos that users explicitly choose, a failure of one underlying does NOT propagate to other markets and is risk the user opted into per-market. Record it under A4 with depth and propagation scope, but do not let it alone drive a red verdict — red requires an external dependency that cross-cuts the protocol or that the user did not opt into at deposit time. A failure mode that is "if the LRT you deposited is hacked, your principal in that LRT-backed market is impaired" is the underlying's autonomy story, not this protocol's; grade it on whether THIS protocol introduces additional dependencies on top.
- Sub-module enumeration is mandatory before grading. If the protocol ships distinct product lines or modules (e.g., a v2 core AMM plus a newer perps/funding-rate module, a lending pool plus a separate vault layer, an L1 core plus a cross-chain extension), enumerate each in findings and grade against the WORST module weighted by its share of TVS or its blast radius. A green core does not rescue an orange/red sub-module; conversely, a small red sub-module with capped TVS may bound the overall grade to orange. Name each module by its on-chain factory or router address. If you do not know whether a module exists, that is an unknowns[] entry, not silence. EXPLICIT TVS WEIGHTING: in the verdict, state each module's approximate share of total protocol TVS (use "~X%" if exact figures unavailable; check DeFiLlama or block-explorer balances on the module's main contract) and how that share informs the weighted grade. Format: "Module A holds ~X% of TVS (grade: <g>); Module B holds ~Y% (grade: <g>); weighted overall = <grade> because <reason>." If a red sub-module holds <5% of total TVS and is capped, the overall grade may be orange; if it holds >25%, the overall grade is red. Do not let qualitative reasoning substitute for the percentages — write the numbers.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. External contract calls. Enumerate every external contract the core contracts call or read from (oracles, price feeds, AMM pools, lending pools, staking/deposit contracts, yield wrappers). For each, identify the address, the provider, and what user-facing function of this protocol would break or mis-price if that external contract paused, mis-reported, or behaved adversarially. Grep for "oracle", "aggregator", "getPrice", "latestAnswer", "chainlink", "pyth", "redstone" as a starting point. To verify an oracle is live and what it currently reports, hit https://defipunkd.com/api/contract/read?chainId=<id>&address=<oracle>&method=latestAnswer (or &method=latestRoundData, &method=getPrice, &method=description; use the BARE method name without parens) and cite the URL — the response includes blockNumber/blockHash and rawReturnData, which is stronger evidence than a docs page about which feed "should" be used.
- A2. Off-chain actor committees reporting INTO the protocol. Oracle committees, guardian multisigs, DAO-selected validator sets acting as protocol reporters, exit-bus signers, fraud-proof challengers. For each, record committee size, quorum, who picks the members, and what mis-reporting could do (mint, burn, finalize withdrawal, freeze). Distinguish this from governance admins (control slice). NOTE ON STAKING PROTOCOLS: validator slashing and node-operator misbehavior are properties of the base-chain substrate, NOT external dependencies, when the operator set is diversified enough that a coordinated failure caps at <5% principal loss. Count validator/operator risk under A2 ONLY if the operator set is small, non-diversified, or lacks bonding / slashing-insurance / diversification mitigations. A curated set of 30+ independent operators with documented diversification falls under the mitigated path; a 3-operator LST does not. Do not cite the protocol's own risk disclosure as evidence that operator failure = principal loss unless you also check the diversification and bond mitigations.
- A3. Bridge / cross-chain messaging dependencies. Only count bridges that carry material TVL or are required for a core user flow. For each, name the bridge operator (canonical L1↔L2, LayerZero, Wormhole, Axelar, custom multisig), the trust model (canonical, optimistic, light-client, guardian set), and what fraction of TVL or users ride on it. "wstETH exists on 15 chains" is not a finding unless material TVL sits there. Before listing any non-primary chain deployment as a dependency, verify it is still operational as of analysis_date — retired or sunset deployments (e.g., Lido-on-Terra, Lido-on-Solana) belong in unknowns[] or should be omitted, never cited as a current dependency.
- A4. Nested collateral / restaking chains. For restaking / LRT / receipt-of-receipt designs: record the depth of the collateral chain, every actor with slashing or freezing power at each level, and whether a failure N levels deep propagates to user principal here.
- A5. Fork lineage (silent check). If DeFiLlama's forkedFrom is non-empty, record it as one finding and move on. If empty, do NOT add a placeholder finding; it adds noise.
- A6. Fallback mechanisms and circuit breakers. What catches an external failure? Sanity-check contracts on oracle reports, rebase bounds, pause paths triggered by bad prices, second-opinion oracles, max-per-block throughput caps, withdrawal queues that absorb bad reports. Record which A1–A4 risks are mitigated by which fallback, and which are unmitigated. For EACH fallback you cite, state its activation status explicitly: (i) LIVE and enforcing on-chain today, (ii) DEPLOYED but not yet wired / activated (e.g., interface exists but the address is zero or the role is unassigned), or (iii) DOCUMENTED / PROPOSED only (forum post, LIP draft, audit pending). Only (i) counts as mitigation for the grade. A fallback in state (ii) or (iii) should be noted but must not reduce the risk in your steel-man or verdict. If you cannot determine activation status, add an unknowns[] entry rather than assume it is live.
- A7. Sequencer / L1-liveness dependency BEYOND the base-chain substrate. SCOPE — sequencer risk only counts here when the protocol IS its own L2/L3 appchain or app-rollup, where the sequencer is part of the protocol's own stack and a freeze is a protocol-level outage. A protocol permissionlessly deployed on a third-party L2 inherits that L2's sequencer as substrate, not an A7 dependency. Record the sequencer/DA trust model when A7 applies.
- A8. Keeper / relayer / off-chain bot liveness. Protocols that need permissionless-but-necessary off-chain actors (liquidation bots, auto-compounders, deposit relayers, intent solvers). Record whether the role is permissionless, what degrades if nobody runs it (yield paused, bad debt accumulates, positions go stale), and whether the failure mode is graceful or catastrophic.
- A9. Governance-mutable dependency surface. Can an admin or DAO action silently INTRODUCE a NEW EXTERNAL dependency — swap the oracle address to a different provider, register a new staking module that calls an untrusted contract, add a new bridge, route SY through a new external vault — without an exit window for users? Check the upgrade / router / module-registry contracts. Answer: which EXTERNAL dependencies are governance-mutable, who holds that power, and whether there is a timelock or exit window. SCOPE LIMIT — read carefully: A9 is about the *external dependency surface*, not the upgrade surface in general. "The proxyAdmin / EOA can upgrade the router implementation to arbitrary bytecode" is a CONTROL-slice finding (admin can rug), NOT an autonomy-A9 finding. A9 fires only when the upgrade specifically swaps out or adds a contract that THIS protocol calls or trusts (e.g., changing the oracle address from Chainlink to a malicious feed, registering a new SY adapter that points to a third-party vault, redirecting a bridge endpoint). If the only finding is "admin can change implementation," do not log it under A9 and do not let it drive the autonomy grade — note it under control instead. The autonomy-relevant version of the same upgrade key is "admin can swap [specific external address X] without timelock"; that requires identifying the specific external dependency that becomes mutable.

STEEL-MAN (per Hard Rule 13): write one-sentence strongest arguments for red, orange, and green using the A1–A9 findings.

IMPACTED TVS ESTIMATE: the headline MUST include a rough impacted-TVS figure — the fraction of protocol TVS that could be lost or frozen if the worst-unmitigated dependency you identified failed. Use "~X%" if exact numbers are unavailable, "<1%" for de minimis, "unclear" only if A1–A9 left the question genuinely open (in which case grade=unknown is usually correct). Do NOT substitute qualitative phrases like "significant" — give a number or bracket.

GRADE ANCHORS (mapped to DefiScan v1 stages):
- green = Stage 2 equivalent. Failure of any external dependency cannot cause loss of user principal or unclaimed yield. Either there are no material external dependencies, or every critical one has a documented fallback (A6) that keeps users whole. Governance cannot silently introduce new dependencies without an exit window (A9). Impacted TVS under any single-dependency failure: effectively 0.
- orange = Stage 1 equivalent. Failure of some external dependency can cause loss of unclaimed yield, or can materially change expected performance (pause withdrawals, freeze positions, degrade price quality), but cannot cause loss of principal. Committee-based oracles with sanity checks, canonical-only bridges, fallback paths that exist but are incomplete, or governance-mutable dependencies protected by a ≥7-day timelock. Impacted TVS is bounded and recoverable.
- red = Stage 0 equivalent. Failure of an external dependency CAN cause theft or loss of principal. Examples: single-provider oracle with no sanity check or fallback, material TVL on a non-canonical bridge with a guardian multisig, governance can hot-swap oracles or add staking modules with no timelock or exit window, unmitigated keeper-liveness dependency where positions become insolvent if bots stop. Impacted TVS is material.
- unknown = checklist incomplete after inspecting source + verified contracts. Prefer unknown over guessing when A1/A6 could not be reconstructed. RESERVE unknown for cases where the CORE ARCHITECTURE itself is unverifiable — not for cases where you merely cannot enumerate every per-market dependency in a multi-market protocol. If the core router/factory/oracle architecture is verifiable on-chain and you can determine whether the core requires external dependencies, grade the architecture even when an exhaustive per-market external-dependency census is infeasible. Acknowledge the per-market gap in unknowns[] but still issue a grade. Refusing to grade a multi-market protocol because you cannot list every SY/vault/market is over-use of unknown; grade the architecture and say so.

PROMPT-META CHECK (per Hard Rule 17): before finalizing, verify the verdict cites concrete contract addresses, docs, or code — not the rubric itself. If your verdict says "the protocol belongs to a category the rubric marks red", rewrite it with the A1–A9 finding that actually justifies the grade, or drop to grade=unknown.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "autonomy",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Open Access Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v1
- protocol.name:              Balancer V1
- protocol.chains:            Ethereum
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://docs.balancer.fi/reference/contracts/security.html#audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: OPEN-ACCESS

Evaluate who is allowed to use the protocol and whether any of that permission is granted off-chain.

Scope: this slice is about ADMISSION — who can enter, exit, or transact. Operator LIVENESS (what breaks if keepers/oracles go offline) is assessed in the dependencies slice and is out of scope for the grade here. You may note operator dependencies as context, but do not let "the protocol halts if operator X disappears" drive the access grade on its own; that belongs in dependencies. Source verification / contract verification on block explorers is assessed in the verifiability slice and is out of scope here — do NOT let "contract is unverified" drive the access grade.

Framing: the smart contracts are the access layer; frontends are UX. A permissionless contract is reachable by any client (SDK, third-party UI, aggregator, wallet integration). Frontend ToS, IP geo-blocking, and wallet screening are publisher policies on one specific client — they are reported as context but do NOT determine the grade. The grade hinges on (1) what the contract itself permits, and (2) whether the protocol is practically reachable without the official publisher's cooperation.

Meta-check before finalizing: if your verdict cites phrases from this prompt as evidence ("the protocol meets the 'credible alternatives' condition", "this fits the 'documented fallback' rule"), redo the verdict. The prompt describes the rubric; evidence must come from the protocol. A verdict should cite what the protocol does, not what the rubric says.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. Whitelist / allowlist modifiers in user-facing entry points. Grep for "onlyWhitelisted", "onlyRole", "allowlist", "isAccredited", "isKYCed". Note which functions are gated and who can add/remove from the list.
- A2. Off-chain operators in the admission path: keepers, sequencers, privileged relayers, oracle posters whose approval is required to admit a user action (not just to keep the protocol live). For each, identify whether the role is held by a single operator, a permissioned committee, or is permissionless. Enumerate per user-facing function class (deposit vs withdraw-request vs claim-finalized vs transfer) which ones require operator approval to be admitted, and which ones admit users unconditionally. A function whose placement is unconditional but whose downstream settlement depends on an operator is an admission-permissionless function — flag the liveness dependency as context and defer its grading weight to the dependencies slice.
- A3. Frontend restrictions on the official interface — record as context, not as a grade lever. Distinguish:
    - A3-passive: boilerplate ToS clauses (sanctions attestation, restricted-territory self-certification, VPN-circumvention prohibition, "comply with applicable law" eligibility, age of majority).
    - A3-active: runtime enforcement — IP-based geo-blocking, wallet-address screening against a sanctions oracle (Chainalysis, TRM, Elliptic), KYC wall, rendering-blocking jurisdiction banner.
  Record findings under the correct tier. Quote ToS text or banner text in evidence[].shows. These findings populate the headline and rationale but do NOT move the grade by themselves; the grade is set by A1, A2, and the A3b path check below.
- A3b. Independent access paths (the operative grade input). Enumerate paths that do not require the official publisher's cooperation:
    - Published SDK / library / CLI for direct contract interaction.
    - Third-party frontends operated by separate legal entities.
    - Wallet-integrated access (MetaMask Swaps, Safe apps, etc.).
    - DEX / lending / yield aggregators that route through the contracts.
  Record at least one concrete link per path that exists. The protocol does NOT have to self-document these — the test is existence, not UX cost. An A3b-i redistribution of the official UI bound by the same ToS does NOT count as an independent path.
- A4. Sanctions / compliance tooling at the contract level: does the protocol check addresses against OFAC lists or similar on-chain blocklists in the contract itself? (Frontend-only screening belongs in A3.)
- A5. Differentiate read access vs write access: many protocols are read-permissionless (anyone can view state) but write-gated (only certain addresses can deposit/borrow). Record both.
- A6. ToS / Legal links: locate them on the website and produce a VERBATIM quote of any jurisdictional, sanctions, or eligibility clause in evidence[].shows. If you cannot extract the clause text verbatim (SPA render failure, paywall, dead link, etc.), do NOT paraphrase or infer from general knowledge — record the ToS URL in unknowns[] with the reason extraction failed. Assertions about ToS content without a verbatim quote will be downweighted by reviewers.

Then write the steel-man section per Hard Rule 11.

Grade rules (admission-focused; liveness concerns belong in dependencies; source verification belongs in verifiability):
- green   = no contract-level whitelist/KYC on user entry/exit; no operator approval required to admit a user action; AND at least one independent A3b path exists (published SDK, third-party frontend, wallet integration, or aggregator routing). Frontend ToS posture and A3-active enforcement on the official UI do NOT block green when contracts are permissionless and an independent path exists — they are reported as context.
- orange  = contracts admit users unconditionally, BUT the protocol is operationally captured by the official publisher: no published SDK, no third-party frontend, no wallet integration, no aggregator routing. The contract is theoretically open but practically reachable only through the official UI. Also applies when admission requires approval from a permissioned committee that is governance-managed with a documented replacement procedure.
- red     = contract-level whitelist / KYC on user entry/exit, OR admission of a core user action requires approval from a single privileged operator or a small committee with no documented replacement procedure, OR enforces an on-chain blocklist updatable by a single party.
- unknown = checklist incomplete after checking the sources above.

Default-grade guidance: when contracts are fully permissionless AND any A3b independent path exists, the default grade is green regardless of frontend ToS or A3-active enforcement on the official UI. Frontend geo-blocking, sanctions-oracle wallet screening, and ToS sanctions clauses are publisher policies on one client and are reported in findings/headline as context, not as grade levers. To grade orange on operational-capture grounds, the auditor must affirmatively show that ALL independent paths are absent or also gated.

Guideline on committees: where admission depends on a multi-operator committee, the relevant axes are (a) set size, (b) whether replacement/rotation is governed on-chain, (c) whether the replacement procedure is publicly documented. A large set with on-chain governance replacement should not be graded as a single-party operator even if rotation is not instantaneous. A small set with informal replacement should be treated as a single-party operator.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "open-access",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Audit all 5 dimensions · one prompt Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v1
- protocol.name:              Balancer V1
- protocol.chains:            Ethereum
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://docs.balancer.fi/reference/contracts/security.html#audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: CONTROL

Evaluate who can change the protocol's rules, how fast, and how broadly.

(Step 0 capability probe and the off-chain-only fallback live in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[])

- **C1.** For each address you assess: who is the contract owner / admin / pendingAdmin / governor — read these via the block explorer's "Read Contract" tab OR `https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=owner` (BARE method names: `&method=owner`, `&method=admin`, `&method=pendingOwner`, `&method=governor`). For Safes use `/api/safe/owners?chainId=<id>&address=0x...`. When a protocol has multiple major versions deployed (v2/v3/v4), perform C1 reads on the NEWEST deployment separately — newer deployments often have weaker control surfaces than the legacy core.
- **C2.** Upgrade mechanism: transparent proxy / UUPS / Beacon / Diamond / immutable. Identify the proxy admin address. Check upgradeability of GOVERNANCE contracts too — a Governor / Aragon Voting / OZ Governor is often itself a proxy whose admin is the Timelock. Asymmetry: when fund-holding cores are immutable AND governance has no admin path that reaches them, an upgradable Governor/Timelock is T3-only and must NOT drag the verdict below green on that basis alone (see grade rules and the "immutable cores" caveat). Only call upgradability "mixed" if you can name a concrete function on the upgradable surface that reaches T1 or T2 on user funds.
- **C3.** EXECUTION PATH (enumerate every stage, in order, with delays in seconds). The operative path is usually a chain (voting → scheduler → timelock → executor; or governor → queue → execute; or Aragon Voting → DualGovernance → EmergencyProtectedTimelock → AdminExecutor). For each stage, record (a) the contract address, (b) the delay constant name + value in seconds, (c) the URL you read it from (block-explorer Read Contract OR `/api/contract/read?...&method=MIN_DELAY&block=<n>`). Do NOT stop at the first timelock-shaped contract — if its admin is itself called by another contract, keep walking. The grading delay is the SUM OF DELAYS ON THE UNCONTESTED FAST PATH (shortest time a proposal with no opposition can go from submission to executable). Dynamic / contested extensions (veto signaling, rage quit, escrow delay) are modifiers, not the basis — note them separately.
- **C4.** Enumerate EVERY multisig with reachable control — main proxy admin, emergency activation, emergency execution, reseal / pause, gate-seal committees, tiebreaker, per-module admins. For each Safe, fetch threshold + owners + version via `/api/safe/owners?chainId=<id>&address=0x...` (response includes raw eth_call data, so the URL is citable evidence). Enumerate ops/council/incentives multisigs even when off the upgrade path — record their scope so a reader can see they are NOT on the upgrade path. For each: (a) address, (b) threshold / total signers, (c) signer identities classified as insider (team, paid auditors under ongoing engagement, mandated service providers) vs non-insider (independent community members, unaffiliated researchers), (d) the specific power held (upgrade, pause, parameter, etc.).
- **C5.** On-chain governance: Governor / GovernorBravo / OZ Governor / Aragon Voting with token-weighted voting? Record proposal threshold, voting period, quorum, and the timelock delay between queue and execute. Every numeric constant must come from a Read Contract call you can link to, or be in unknowns[] with the C-code. If votingDelay / votingPeriod are denominated in BLOCKS, convert to seconds at the chain's CURRENT block time (Ethereum mainnet ≈ 12s post-Merge, not the 15s in older Compound/Bravo deployments) — cite both block count and converted seconds.
- **C6.** EMERGENCY POWERS: separate emergency-pause / guardian role with a different time cap or different actor than the main upgrade authority? Record it explicitly.
- **C7.** POWER TIER (blast radius). For each privileged path in C3–C6, classify the WORST thing that path can do, choosing the highest applicable tier. Cite the specific function name and any on-chain bound — tier claims without a named function are unsupported.
    - **T1 — FUND-CRITICAL**: replace implementation of contracts holding user funds; change AMM math / accounting / collateral logic; mint unbacked debt or shares; pause withdrawals; drain user-fund treasury; change oracle to attacker-controlled source; replace upgrade admin with EOA.
    - **T2 — ECONOMICALLY MATERIAL**: change fee parameters within bounded ranges; redirect protocol fees; add/remove markets / collateral types; bounded inflation or token mint within hard-capped schedule; spend protocol-owned (non-user) treasury.
    - **T3 — GOVERNANCE-INTERNAL**: change voting rules, quorum, voting period, proposal threshold; upgrade the Governor itself; rotate Timelock admin; mint governance tokens within a capped annual schedule.
    - **T4 — OPERATIONAL**: incentives distribution, grants, ENS / frontend canonicalization, deployment coordination, periphery router deprecation.
  The grade is set by the HIGHEST tier reachable on the uncontested fast path, not the median. State the tier and the binding function in the verdict.

### Read Contract discipline (applies to C3, C4, C5)

Every numeric constant cited (timelock delays, voting periods, multisig thresholds, quorum percentages) must come from EITHER (a) a block-explorer Read Contract URL, OR (b) a DeFiPunkd `/api/contract/read` or `/api/safe/owners` URL (preferred with `&block=<n>` for content-addressed evidence), OR (c) an unknowns[] entry with the C-code. Docs / blog posts are corroboration only — they cannot be the sole citation for a value that is also readable on-chain.

### Off-chain-only substitute hierarchy (when grading_basis="off-chain-only" — see preamble Rule 16)

When on-chain reads were genuinely unreachable this run, eligible off-chain substitutes in priority order:
1. Linked audit PDFs (admin roles, multisig members, timelock delays usually enumerated).
2. Governance forum posts that quote constants from a successful on-chain proposal (cite post URL + linked execution-tx URL).
3. Official protocol docs pages with named addresses and roles (must be on a domain owned by the protocol).
4. GitHub README / SECURITY.md / governance/*.md at a pinned commit SHA.

Forbidden substitutes: third-party blog posts, X / Twitter threads, search-result snippets, model memory. Required degradation: any C-code citing a numeric constant from docs/forum/audit prose ONLY must also carry an `unknowns[]` entry with `-offchain` suffix noting "value not re-read on-chain in this run; corroboration only".

### Grade rules (apply the timelock bar conditional on the highest C7 tier reachable on the fast path)

Security Council standard (used below): a multisig qualifies as "Security Council" only if ALL of: ≥7 signers, ≥51% threshold, ≥50% non-insider signers, every signer publicly announced. Failing any criterion = NOT a Security Council, regardless of signer reputation.

- **green**: highest reachable tier is T3 or T4 regardless of timelock; OR T2 reachable with uncontested-fast-path delay ≥7 days; OR T1 reachable only via immutable contracts (T1 is unreachable); OR T1 reachable with uncontested-fast-path delay ≥7 days combined with a Security Council multisig; OR T1 reachable with uncontested-fast-path delay ≥7 days through active on-chain governance with broad token distribution.
- **orange**: T2 reachable with uncontested-fast-path delay >0 but <7 days; OR T1 reachable with uncontested-fast-path delay >0 but <7 days; OR a multisig failing one or more Security Council criteria sits on a T1/T2 path; OR unclear upgrade authority on a T1/T2 path; OR governance with very short timelock or low quorum on a T1/T2 path.
- **red**: T1 reachable with no timelock by a single EOA or 2-of-3 multisig; OR a T1 upgrade admin that is not a smart contract you can audit.
- **unknown**: completed the checklist but still cannot determine the upgrade authority OR cannot classify the highest tier reachable on the main contracts.

Tiering caveats:
- "Bounded" must be enforced ON-CHAIN to count as T2. A function that sets fees with no upper-bound check is T1 — cite the bound check.
- Recurring T2 economic extraction (e.g. fee redirect with no rate limit) approaches T1 over time. A single proposal that can permanently redirect all future revenue is T1.
- T3 assumes the governance contract cannot itself authorize a T1/T2 action without going through the same timelock. If governance can self-upgrade to bypass the timelock, T3 collapses into T1.
- Do not downgrade tier by hand-waving ("realistically governance would never…"). Tier on what the contract permits, not what feels likely.

Notes:
- **Dynamic / dual-governance timelocks** (Lido, Compound escrow veto): the rubric grades on the uncontested path because that is the path most upgrades take. A dynamic extension that fires only under stake-weighted opposition is a real protection — name it in the green steel-man, but it does not lift an orange fast path into green; state the tension in the verdict.
- **Immutable cores with upgradable governance** (Uniswap-style): if fund-holding contracts are immutable and have no admin-reachable function moving / freezing / re-routing user funds, the highest reachable tier on the upgrade path is T3 — green regardless of timelock. Don't grade this orange just because the Governor is a proxy — that's a C2 fact, not a downgrade. Downgrade only applies if you can cite a concrete function on the upgradable surface that reaches T1 or T2 (privileged hook, upgradable factory controlling fund-routing, fee-switch redirecting protocol revenue without bound).
- **The 7-day bar** reflects the exit-window standard — users need notice after a queued upgrade to withdraw if they disagree. The ability-to-exit slice grades the exit side; this slice grades the delay side; both must hold for users to actually benefit from the delay.

---

### Slice: ABILITY-TO-EXIT

Evaluate whether users can withdraw their funds on their own terms, even under adversarial admin conditions.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- E1. Enumerate every user-facing exit function in the main contracts: withdraw, redeem, burn, requestWithdrawal, claim, exit, etc. List them by name. Do NOT treat the contract as a monolith.
- E2. For EACH exit function in E1: identify its access modifiers and any pause guards (e.g. _checkResumed, whenNotPaused, onlyRole). Functions that gate REQUEST PLACEMENT often differ from functions that CLAIM FINALIZED FUNDS — check both separately.
- E3. For each pause guard: identify the role holder (which address holds PAUSE_ROLE / GUARDIAN / etc.) and the maximum pause duration. Specifically check whether PAUSE_INFINITELY (or equivalent uncapped pause) is callable, and which actor can call it (single multisig vs governance vote). For role-holder reads use https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=hasRole&args=0x...,0x... or &method=getRoleAdmin&args=0x.... For "is currently paused" checks use &method=paused or &method=isPaused&args=<resume-code>. Use the BARE method name (no parens). Cite the URL with &block=<n> in evidence[].
- E4. EMERGENCY vs GOVERNANCE pause distinction: many protocols have a fast-acting emergency pause capped at N days and a slow governance pause that can be indefinite. Record both paths separately if present, with their time caps and actor classes.
- E5. Queued redemption: documented maximum queue duration, daily withdrawal caps, whether the queue itself is pausable.
- E6. Forced-exit / escape-hatch / permissionless emergency-exit mechanism for adversarial-admin scenarios.
- E7. Frontend dependency: confirm exit functions are directly callable on-chain (e.g. via Etherscan write tab or a generic wallet) without the project's frontend.

Then write the steel-man section per Hard Rule 11. Common red-vs-orange tension on this slice: indefinite pause exists (suggests red) BUT the realistic emergency path is time-capped AND claims of already-finalized exits are not pause-gated (suggests orange). Resolve this by stating who can do what for how long, not by stopping at the worst-case sentence.

Grade rules:
- green   = permissionless exit; pause is either absent, narrowly scoped to clearly-described emergencies with auto-expiry, or capped at ≤7 days; no frontend dependency for exit; claims of already-finalized exits are not pause-gated under any path.
- orange  = pausable with broad scope OR indefinite pause is reachable only through governance vote (not unilateral admin action), OR queued redemption with documented max > 7 days, OR claims-of-finalized are exempt but new-request placement can be paused indefinitely by governance.
- red     = exit requires admin signature, OR ANY actor (including governance) can pause CLAIMS of finalized exits indefinitely, OR there is no on-chain exit function at all (purely custodial), OR pause is held by a single EOA / 2-of-3 multisig with no time cap.
- unknown = checklist incomplete after checking the sources above.

---

### Slice: AUTONOMY

Evaluate this protocol's autonomy: can a failure of anything outside its own contracts cause theft or loss of user principal, loss of unclaimed yield, or materially change the protocol's expected performance? "Autonomy" is not "has zero external touchpoints" — it is "external failures are either survivable or recoverable without user loss". This slice adapts DefiScan v1's Autonomy dimension; dependencies are one of several criteria, not the whole frame.

GUARDRAILS (read before grading):
- Category alone (Liquid Staking, Bridge, RWA Lending, Restaking, …) does NOT force a grade. A category is a hint about where to look; the grade must come from the concrete A1–A9 findings below.
- Base-chain consensus (Ethereum PoS, the chain's validator set, the canonical Deposit Contract at 0x00000000219ab540356cBB839Cbe05303d7705Fa) is the SUBSTRATE, not a dependency, for any protocol deployed on that chain. Do not list "depends on Ethereum" as a finding.
- Oracles or other integrations used by DOWNSTREAM protocols that happen to read this protocol's token (e.g. Chainlink stETH/USD consumed by Aave) are NOT this protocol's dependencies. Count only what THIS protocol's contracts call or trust on-chain. If a feed or contract appears in the protocol's docs only as reference material for third-party integrators, EXCLUDE IT ENTIRELY — do not log it as a finding even with a "peripheral" or "referenced only" caveat.
- "Upgradeable admin can change things" belongs to the CONTROL slice; only count it here when the upgrade surface lets governance silently swap an external dependency (see A9).
- Underlying-asset risk in opt-in, isolated markets is NOT autonomy-red on its own. When a protocol wraps third-party yield-bearing assets (LSTs, LRTs, ERC-4626 vaults, lending receipts, restaked tokens) into per-market silos that users explicitly choose, a failure of one underlying does NOT propagate to other markets and is risk the user opted into per-market. Record it under A4 with depth and propagation scope, but do not let it alone drive a red verdict — red requires an external dependency that cross-cuts the protocol or that the user did not opt into at deposit time. A failure mode that is "if the LRT you deposited is hacked, your principal in that LRT-backed market is impaired" is the underlying's autonomy story, not this protocol's; grade it on whether THIS protocol introduces additional dependencies on top.
- Sub-module enumeration is mandatory before grading. If the protocol ships distinct product lines or modules (e.g., a v2 core AMM plus a newer perps/funding-rate module, a lending pool plus a separate vault layer, an L1 core plus a cross-chain extension), enumerate each in findings and grade against the WORST module weighted by its share of TVS or its blast radius. A green core does not rescue an orange/red sub-module; conversely, a small red sub-module with capped TVS may bound the overall grade to orange. Name each module by its on-chain factory or router address. If you do not know whether a module exists, that is an unknowns[] entry, not silence. EXPLICIT TVS WEIGHTING: in the verdict, state each module's approximate share of total protocol TVS (use "~X%" if exact figures unavailable; check DeFiLlama or block-explorer balances on the module's main contract) and how that share informs the weighted grade. Format: "Module A holds ~X% of TVS (grade: <g>); Module B holds ~Y% (grade: <g>); weighted overall = <grade> because <reason>." If a red sub-module holds <5% of total TVS and is capped, the overall grade may be orange; if it holds >25%, the overall grade is red. Do not let qualitative reasoning substitute for the percentages — write the numbers.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. External contract calls. Enumerate every external contract the core contracts call or read from (oracles, price feeds, AMM pools, lending pools, staking/deposit contracts, yield wrappers). For each, identify the address, the provider, and what user-facing function of this protocol would break or mis-price if that external contract paused, mis-reported, or behaved adversarially. Grep for "oracle", "aggregator", "getPrice", "latestAnswer", "chainlink", "pyth", "redstone" as a starting point. To verify an oracle is live and what it currently reports, hit https://defipunkd.com/api/contract/read?chainId=<id>&address=<oracle>&method=latestAnswer (or &method=latestRoundData, &method=getPrice, &method=description; use the BARE method name without parens) and cite the URL — the response includes blockNumber/blockHash and rawReturnData, which is stronger evidence than a docs page about which feed "should" be used.
- A2. Off-chain actor committees reporting INTO the protocol. Oracle committees, guardian multisigs, DAO-selected validator sets acting as protocol reporters, exit-bus signers, fraud-proof challengers. For each, record committee size, quorum, who picks the members, and what mis-reporting could do (mint, burn, finalize withdrawal, freeze). Distinguish this from governance admins (control slice). NOTE ON STAKING PROTOCOLS: validator slashing and node-operator misbehavior are properties of the base-chain substrate, NOT external dependencies, when the operator set is diversified enough that a coordinated failure caps at <5% principal loss. Count validator/operator risk under A2 ONLY if the operator set is small, non-diversified, or lacks bonding / slashing-insurance / diversification mitigations. A curated set of 30+ independent operators with documented diversification falls under the mitigated path; a 3-operator LST does not. Do not cite the protocol's own risk disclosure as evidence that operator failure = principal loss unless you also check the diversification and bond mitigations.
- A3. Bridge / cross-chain messaging dependencies. Only count bridges that carry material TVL or are required for a core user flow. For each, name the bridge operator (canonical L1↔L2, LayerZero, Wormhole, Axelar, custom multisig), the trust model (canonical, optimistic, light-client, guardian set), and what fraction of TVL or users ride on it. "wstETH exists on 15 chains" is not a finding unless material TVL sits there. Before listing any non-primary chain deployment as a dependency, verify it is still operational as of analysis_date — retired or sunset deployments (e.g., Lido-on-Terra, Lido-on-Solana) belong in unknowns[] or should be omitted, never cited as a current dependency.
- A4. Nested collateral / restaking chains. For restaking / LRT / receipt-of-receipt designs: record the depth of the collateral chain, every actor with slashing or freezing power at each level, and whether a failure N levels deep propagates to user principal here.
- A5. Fork lineage (silent check). If DeFiLlama's forkedFrom is non-empty, record it as one finding and move on. If empty, do NOT add a placeholder finding; it adds noise.
- A6. Fallback mechanisms and circuit breakers. What catches an external failure? Sanity-check contracts on oracle reports, rebase bounds, pause paths triggered by bad prices, second-opinion oracles, max-per-block throughput caps, withdrawal queues that absorb bad reports. Record which A1–A4 risks are mitigated by which fallback, and which are unmitigated. For EACH fallback you cite, state its activation status explicitly: (i) LIVE and enforcing on-chain today, (ii) DEPLOYED but not yet wired / activated (e.g., interface exists but the address is zero or the role is unassigned), or (iii) DOCUMENTED / PROPOSED only (forum post, LIP draft, audit pending). Only (i) counts as mitigation for the grade. A fallback in state (ii) or (iii) should be noted but must not reduce the risk in your steel-man or verdict. If you cannot determine activation status, add an unknowns[] entry rather than assume it is live.
- A7. Sequencer / L1-liveness dependency BEYOND the base-chain substrate. SCOPE — sequencer risk only counts here when the protocol IS its own L2/L3 appchain or app-rollup, where the sequencer is part of the protocol's own stack and a freeze is a protocol-level outage. A protocol permissionlessly deployed on a third-party L2 inherits that L2's sequencer as substrate, not an A7 dependency. Record the sequencer/DA trust model when A7 applies.
- A8. Keeper / relayer / off-chain bot liveness. Protocols that need permissionless-but-necessary off-chain actors (liquidation bots, auto-compounders, deposit relayers, intent solvers). Record whether the role is permissionless, what degrades if nobody runs it (yield paused, bad debt accumulates, positions go stale), and whether the failure mode is graceful or catastrophic.
- A9. Governance-mutable dependency surface. Can an admin or DAO action silently INTRODUCE a NEW EXTERNAL dependency — swap the oracle address to a different provider, register a new staking module that calls an untrusted contract, add a new bridge, route SY through a new external vault — without an exit window for users? Check the upgrade / router / module-registry contracts. Answer: which EXTERNAL dependencies are governance-mutable, who holds that power, and whether there is a timelock or exit window. SCOPE LIMIT — read carefully: A9 is about the *external dependency surface*, not the upgrade surface in general. "The proxyAdmin / EOA can upgrade the router implementation to arbitrary bytecode" is a CONTROL-slice finding (admin can rug), NOT an autonomy-A9 finding. A9 fires only when the upgrade specifically swaps out or adds a contract that THIS protocol calls or trusts (e.g., changing the oracle address from Chainlink to a malicious feed, registering a new SY adapter that points to a third-party vault, redirecting a bridge endpoint). If the only finding is "admin can change implementation," do not log it under A9 and do not let it drive the autonomy grade — note it under control instead. The autonomy-relevant version of the same upgrade key is "admin can swap [specific external address X] without timelock"; that requires identifying the specific external dependency that becomes mutable.

STEEL-MAN (per Hard Rule 13): write one-sentence strongest arguments for red, orange, and green using the A1–A9 findings.

IMPACTED TVS ESTIMATE: the headline MUST include a rough impacted-TVS figure — the fraction of protocol TVS that could be lost or frozen if the worst-unmitigated dependency you identified failed. Use "~X%" if exact numbers are unavailable, "<1%" for de minimis, "unclear" only if A1–A9 left the question genuinely open (in which case grade=unknown is usually correct). Do NOT substitute qualitative phrases like "significant" — give a number or bracket.

GRADE ANCHORS (mapped to DefiScan v1 stages):
- green = Stage 2 equivalent. Failure of any external dependency cannot cause loss of user principal or unclaimed yield. Either there are no material external dependencies, or every critical one has a documented fallback (A6) that keeps users whole. Governance cannot silently introduce new dependencies without an exit window (A9). Impacted TVS under any single-dependency failure: effectively 0.
- orange = Stage 1 equivalent. Failure of some external dependency can cause loss of unclaimed yield, or can materially change expected performance (pause withdrawals, freeze positions, degrade price quality), but cannot cause loss of principal. Committee-based oracles with sanity checks, canonical-only bridges, fallback paths that exist but are incomplete, or governance-mutable dependencies protected by a ≥7-day timelock. Impacted TVS is bounded and recoverable.
- red = Stage 0 equivalent. Failure of an external dependency CAN cause theft or loss of principal. Examples: single-provider oracle with no sanity check or fallback, material TVL on a non-canonical bridge with a guardian multisig, governance can hot-swap oracles or add staking modules with no timelock or exit window, unmitigated keeper-liveness dependency where positions become insolvent if bots stop. Impacted TVS is material.
- unknown = checklist incomplete after inspecting source + verified contracts. Prefer unknown over guessing when A1/A6 could not be reconstructed. RESERVE unknown for cases where the CORE ARCHITECTURE itself is unverifiable — not for cases where you merely cannot enumerate every per-market dependency in a multi-market protocol. If the core router/factory/oracle architecture is verifiable on-chain and you can determine whether the core requires external dependencies, grade the architecture even when an exhaustive per-market external-dependency census is infeasible. Acknowledge the per-market gap in unknowns[] but still issue a grade. Refusing to grade a multi-market protocol because you cannot list every SY/vault/market is over-use of unknown; grade the architecture and say so.

PROMPT-META CHECK (per Hard Rule 17): before finalizing, verify the verdict cites concrete contract addresses, docs, or code — not the rubric itself. If your verdict says "the protocol belongs to a category the rubric marks red", rewrite it with the A1–A9 finding that actually justifies the grade, or drop to grade=unknown.

---

### Slice: OPEN-ACCESS

Evaluate who is allowed to use the protocol and whether any of that permission is granted off-chain.

Scope: this slice is about ADMISSION — who can enter, exit, or transact. Operator LIVENESS (what breaks if keepers/oracles go offline) is assessed in the dependencies slice and is out of scope for the grade here. You may note operator dependencies as context, but do not let "the protocol halts if operator X disappears" drive the access grade on its own; that belongs in dependencies. Source verification / contract verification on block explorers is assessed in the verifiability slice and is out of scope here — do NOT let "contract is unverified" drive the access grade.

Framing: the smart contracts are the access layer; frontends are UX. A permissionless contract is reachable by any client (SDK, third-party UI, aggregator, wallet integration). Frontend ToS, IP geo-blocking, and wallet screening are publisher policies on one specific client — they are reported as context but do NOT determine the grade. The grade hinges on (1) what the contract itself permits, and (2) whether the protocol is practically reachable without the official publisher's cooperation.

Meta-check before finalizing: if your verdict cites phrases from this prompt as evidence ("the protocol meets the 'credible alternatives' condition", "this fits the 'documented fallback' rule"), redo the verdict. The prompt describes the rubric; evidence must come from the protocol. A verdict should cite what the protocol does, not what the rubric says.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. Whitelist / allowlist modifiers in user-facing entry points. Grep for "onlyWhitelisted", "onlyRole", "allowlist", "isAccredited", "isKYCed". Note which functions are gated and who can add/remove from the list.
- A2. Off-chain operators in the admission path: keepers, sequencers, privileged relayers, oracle posters whose approval is required to admit a user action (not just to keep the protocol live). For each, identify whether the role is held by a single operator, a permissioned committee, or is permissionless. Enumerate per user-facing function class (deposit vs withdraw-request vs claim-finalized vs transfer) which ones require operator approval to be admitted, and which ones admit users unconditionally. A function whose placement is unconditional but whose downstream settlement depends on an operator is an admission-permissionless function — flag the liveness dependency as context and defer its grading weight to the dependencies slice.
- A3. Frontend restrictions on the official interface — record as context, not as a grade lever. Distinguish:
    - A3-passive: boilerplate ToS clauses (sanctions attestation, restricted-territory self-certification, VPN-circumvention prohibition, "comply with applicable law" eligibility, age of majority).
    - A3-active: runtime enforcement — IP-based geo-blocking, wallet-address screening against a sanctions oracle (Chainalysis, TRM, Elliptic), KYC wall, rendering-blocking jurisdiction banner.
  Record findings under the correct tier. Quote ToS text or banner text in evidence[].shows. These findings populate the headline and rationale but do NOT move the grade by themselves; the grade is set by A1, A2, and the A3b path check below.
- A3b. Independent access paths (the operative grade input). Enumerate paths that do not require the official publisher's cooperation:
    - Published SDK / library / CLI for direct contract interaction.
    - Third-party frontends operated by separate legal entities.
    - Wallet-integrated access (MetaMask Swaps, Safe apps, etc.).
    - DEX / lending / yield aggregators that route through the contracts.
  Record at least one concrete link per path that exists. The protocol does NOT have to self-document these — the test is existence, not UX cost. An A3b-i redistribution of the official UI bound by the same ToS does NOT count as an independent path.
- A4. Sanctions / compliance tooling at the contract level: does the protocol check addresses against OFAC lists or similar on-chain blocklists in the contract itself? (Frontend-only screening belongs in A3.)
- A5. Differentiate read access vs write access: many protocols are read-permissionless (anyone can view state) but write-gated (only certain addresses can deposit/borrow). Record both.
- A6. ToS / Legal links: locate them on the website and produce a VERBATIM quote of any jurisdictional, sanctions, or eligibility clause in evidence[].shows. If you cannot extract the clause text verbatim (SPA render failure, paywall, dead link, etc.), do NOT paraphrase or infer from general knowledge — record the ToS URL in unknowns[] with the reason extraction failed. Assertions about ToS content without a verbatim quote will be downweighted by reviewers.

Then write the steel-man section per Hard Rule 11.

Grade rules (admission-focused; liveness concerns belong in dependencies; source verification belongs in verifiability):
- green   = no contract-level whitelist/KYC on user entry/exit; no operator approval required to admit a user action; AND at least one independent A3b path exists (published SDK, third-party frontend, wallet integration, or aggregator routing). Frontend ToS posture and A3-active enforcement on the official UI do NOT block green when contracts are permissionless and an independent path exists — they are reported as context.
- orange  = contracts admit users unconditionally, BUT the protocol is operationally captured by the official publisher: no published SDK, no third-party frontend, no wallet integration, no aggregator routing. The contract is theoretically open but practically reachable only through the official UI. Also applies when admission requires approval from a permissioned committee that is governance-managed with a documented replacement procedure.
- red     = contract-level whitelist / KYC on user entry/exit, OR admission of a core user action requires approval from a single privileged operator or a small committee with no documented replacement procedure, OR enforces an on-chain blocklist updatable by a single party.
- unknown = checklist incomplete after checking the sources above.

Default-grade guidance: when contracts are fully permissionless AND any A3b independent path exists, the default grade is green regardless of frontend ToS or A3-active enforcement on the official UI. Frontend geo-blocking, sanctions-oracle wallet screening, and ToS sanctions clauses are publisher policies on one client and are reported in findings/headline as context, not as grade levers. To grade orange on operational-capture grounds, the auditor must affirmatively show that ALL independent paths are absent or also gated.

Guideline on committees: where admission depends on a multi-operator committee, the relevant axes are (a) set size, (b) whether replacement/rotation is governed on-chain, (c) whether the replacement procedure is publicly documented. A large set with on-chain governance replacement should not be graded as a single-party operator even if rotation is not instantaneous. A small set with informal replacement should be treated as a single-party operator.

---

### Slice: VERIFIABILITY

Evaluate whether an outsider can independently confirm what the deployed code does.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- V1. For each address you assess: is the bytecode verified on the chain's block explorer? Record the "Contract Source Code Verified" indicator. https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x... returns this as a top-level "verified" boolean plus "abiSource" ("etherscan" / "sourcify") and an inline ABI — useful when the explorer page is rate-limited. If the contract is a proxy, verify BOTH the proxy contract AND the current implementation contract. The same /api/contract/abi response auto-resolves proxies and includes a "proxy.implementation" address when present, so one call covers V1 + V6 in one shot. An explorer "Similar Match" on a well-known proxy pattern (Aragon AppProxyUpgradeable, ERC1967Proxy, OssifiableProxy, OZ TransparentUpgradeableProxy) is expected for that pattern and does NOT count as a verification gap on its own — what matters is that the implementation is independently verified.
- V2. Source-to-repo correspondence: for each verified contract, attempt to find a matching commit in the linked GitHub repos. Record evidence[].commit on a match. Independent compile/bytecode-match is NOT required for green — a recognized public repo whose structure and file contents correspond to the explorer-visible source is sufficient. If you did not pin a commit SHA or run a bytecode diff, record that plainly in unknowns[] and proceed; it is a scope limit, not a downgrade signal.
- V3. Audit coverage: for each URL in protocol.audit_links, open it and record: auditor name, audit date, the specific contracts / commit in scope. Flag audits that predate the current deployment by >6 months without a follow-up review.
- V4. Auditor recognition: the following firms are broadly recognized in Solidity: Trail of Bits, Zellic, Spearbit, OpenZeppelin, ConsenSys Diligence, Certora (formal verification), Quantstamp, Halborn, Peckshield, Sigma Prime, ChainSecurity, Ackee Blockchain, MixBytes, Statemind. Unknown firms are orange-at-best for any green-grade claim. Name the firm explicitly in evidence[].
- V5. Post-audit drift: compare the most recent in-scope audit(s) against the currently-deployed source, weighted by what each contract does and by what the changes actually contain. SCOPE — drift only downgrades the grade when ALL of the following hold: (i) the drifting contract is fund-custody / settlement / accounting-critical (NOT a peripheral router, lens, quoter, or pure-view contract that holds no balances); (ii) the changes are material — new functions, modified access control, modified accounting, modified fund flow — and not refactors / struct relocations / import reorgs / build or CI fixes / formatting; (iii) no later audit, fix-audit, or differential audit from a recognized firm covers the changed files (audits often pin a pre-fix commit while a follow-up reviews the delta — match by file scope, not by commit-hash equality). When you cite drift as a downgrade reason, name the specific behavior change (function added, role granted, accounting formula altered) — "N commits ahead" or "+X/-Y LOC" alone is not evidence of material drift. If you have not sampled the diff content (e.g. via the GitHub compare view or the top commits in the window), record drift as an unknowns[] entry rather than auto-downgrading; commit-count and LOC are starting signals, not findings.
- V6. Implementation vs proxy: a verified proxy with an unverified implementation is effectively unverified. State whether the implementation is verified separately.

EVIDENCE DISCIPLINE (read before writing findings[]):
- Do not assert a specific deploy-commit SHA, bytecode equivalence, or "identical to audited commit" unless you actually fetched the artifact that shows it (e.g., a deployed-addresses JSON you opened, an explorer page you read). Inferred or plausible matches belong in unknowns[], never in findings[] or evidence[].
- Evidence[] entries must correspond to pages/files you actually retrieved this run. A URL you did not open is not evidence.

Then write the steel-man section per Hard Rule 11.

Grade rules:
- green   = deployed bytecode verified on the explorer (proxy AND implementation if proxied; "Similar Match" on a standard proxy pattern is fine per V1), a public source repo exists whose contents correspond to the explorer-visible source, AND ≥1 audit from a recognized firm covering the currently-deployed contracts (≤6 months of drift OR drift was re-audited). A missing local compile-match is not a downgrade — record it in unknowns[] and still grade green if the other conditions hold.
- orange  = verified but with visible drift from the public repo, OR audit scope is stale relative to deployment, OR only minor / unknown-firm audits exist, OR only some of the main contracts are verified, OR proxy verified but implementation only partially verified.
- red     = unverified bytecode (or verified proxy with unverified implementation), OR no audit in protocol.audit_links, OR no public repo.
- unknown = reserved for when the protocol's verifiability posture genuinely cannot be assessed (e.g., explorer and repo both inaccessible for this protocol). Do NOT use unknown merely because you, the analyst, could not run a particular check such as a bytecode diff — that goes in unknowns[] while the grade is still assigned from the evidence you do have.

---

### JSON output contract

Return exactly one JSON array inside a single ```json fenced block. The array MUST contain exactly five objects, one per risk slice, in this exact order: "control", "ability-to-exit", "autonomy", "open-access", "verifiability". Do not include the discovery slice.

Each object has the same shape as a normal slice submission:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "<one of: control | ability-to-exit | autonomy | open-access | verifiability>",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- Produce one complete object for each of these slices only: control, ability-to-exit, autonomy, open-access, verifiability.
- Reuse the same model, chat_url, snapshot_generated_at, prompt_version, analysis_date, and slug values across all five objects.
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches that object's slice checklist prefix verbatim (C1, E2, AU3, A3b, V4a, …); unknowns[] entries are checklist-coded ("C3: …").
- Wrap the array in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Verifiability tentative

Open source + 3 audits

Protocol publishes a GitHub repository and has at least one audit on record. This is a coarse Phase-0 signal only: auditor reputation, scope, and post-audit review coverage are not yet weighted.

Run your own prompt Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v1
- protocol.name:              Balancer V1
- protocol.chains:            Ethereum
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://docs.balancer.fi/reference/contracts/security.html#audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: VERIFIABILITY

Evaluate whether an outsider can independently confirm what the deployed code does.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- V1. For each address you assess: is the bytecode verified on the chain's block explorer? Record the "Contract Source Code Verified" indicator. https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x... returns this as a top-level "verified" boolean plus "abiSource" ("etherscan" / "sourcify") and an inline ABI — useful when the explorer page is rate-limited. If the contract is a proxy, verify BOTH the proxy contract AND the current implementation contract. The same /api/contract/abi response auto-resolves proxies and includes a "proxy.implementation" address when present, so one call covers V1 + V6 in one shot. An explorer "Similar Match" on a well-known proxy pattern (Aragon AppProxyUpgradeable, ERC1967Proxy, OssifiableProxy, OZ TransparentUpgradeableProxy) is expected for that pattern and does NOT count as a verification gap on its own — what matters is that the implementation is independently verified.
- V2. Source-to-repo correspondence: for each verified contract, attempt to find a matching commit in the linked GitHub repos. Record evidence[].commit on a match. Independent compile/bytecode-match is NOT required for green — a recognized public repo whose structure and file contents correspond to the explorer-visible source is sufficient. If you did not pin a commit SHA or run a bytecode diff, record that plainly in unknowns[] and proceed; it is a scope limit, not a downgrade signal.
- V3. Audit coverage: for each URL in protocol.audit_links, open it and record: auditor name, audit date, the specific contracts / commit in scope. Flag audits that predate the current deployment by >6 months without a follow-up review.
- V4. Auditor recognition: the following firms are broadly recognized in Solidity: Trail of Bits, Zellic, Spearbit, OpenZeppelin, ConsenSys Diligence, Certora (formal verification), Quantstamp, Halborn, Peckshield, Sigma Prime, ChainSecurity, Ackee Blockchain, MixBytes, Statemind. Unknown firms are orange-at-best for any green-grade claim. Name the firm explicitly in evidence[].
- V5. Post-audit drift: compare the most recent in-scope audit(s) against the currently-deployed source, weighted by what each contract does and by what the changes actually contain. SCOPE — drift only downgrades the grade when ALL of the following hold: (i) the drifting contract is fund-custody / settlement / accounting-critical (NOT a peripheral router, lens, quoter, or pure-view contract that holds no balances); (ii) the changes are material — new functions, modified access control, modified accounting, modified fund flow — and not refactors / struct relocations / import reorgs / build or CI fixes / formatting; (iii) no later audit, fix-audit, or differential audit from a recognized firm covers the changed files (audits often pin a pre-fix commit while a follow-up reviews the delta — match by file scope, not by commit-hash equality). When you cite drift as a downgrade reason, name the specific behavior change (function added, role granted, accounting formula altered) — "N commits ahead" or "+X/-Y LOC" alone is not evidence of material drift. If you have not sampled the diff content (e.g. via the GitHub compare view or the top commits in the window), record drift as an unknowns[] entry rather than auto-downgrading; commit-count and LOC are starting signals, not findings.
- V6. Implementation vs proxy: a verified proxy with an unverified implementation is effectively unverified. State whether the implementation is verified separately.

EVIDENCE DISCIPLINE (read before writing findings[]):
- Do not assert a specific deploy-commit SHA, bytecode equivalence, or "identical to audited commit" unless you actually fetched the artifact that shows it (e.g., a deployed-addresses JSON you opened, an explorer page you read). Inferred or plausible matches belong in unknowns[], never in findings[] or evidence[].
- Evidence[] entries must correspond to pages/files you actually retrieved this run. A URL you did not open is not evidence.

Then write the steel-man section per Hard Rule 11.

Grade rules:
- green   = deployed bytecode verified on the explorer (proxy AND implementation if proxied; "Similar Match" on a standard proxy pattern is fine per V1), a public source repo exists whose contents correspond to the explorer-visible source, AND ≥1 audit from a recognized firm covering the currently-deployed contracts (≤6 months of drift OR drift was re-audited). A missing local compile-match is not a downgrade — record it in unknowns[] and still grade green if the other conditions hold.
- orange  = verified but with visible drift from the public repo, OR audit scope is stale relative to deployment, OR only minor / unknown-firm audits exist, OR only some of the main contracts are verified, OR proxy verified but implementation only partially verified.
- red     = unverified bytecode (or verified proxy with unverified implementation), OR no audit in protocol.audit_links, OR no public repo.
- unknown = reserved for when the protocol's verifiability posture genuinely cannot be assessed (e.g., explorer and repo both inaccessible for this protocol). Do NOT use unknown merely because you, the analyst, could not run a particular check such as a bytecode diff — that goes in unknowns[] while the grade is still assigned from the evidence you do have.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "verifiability",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

4 dimensions not yet assessed (Control, Ability to exit, Autonomy, Open Access)

Control unknown Unverified

Not yet assessed

No model has graded this dimension yet. Run the slice prompt through any LLM and submit the JSON — once ≥3 independent runs agree, the quorum bot merges the verdict here.

Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v1
- protocol.name:              Balancer V1
- protocol.chains:            Ethereum
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://docs.balancer.fi/reference/contracts/security.html#audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: CONTROL

Evaluate who can change the protocol's rules, how fast, and how broadly.

(Step 0 capability probe and the off-chain-only fallback live in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[])

- **C1.** For each address you assess: who is the contract owner / admin / pendingAdmin / governor — read these via the block explorer's "Read Contract" tab OR `https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=owner` (BARE method names: `&method=owner`, `&method=admin`, `&method=pendingOwner`, `&method=governor`). For Safes use `/api/safe/owners?chainId=<id>&address=0x...`. When a protocol has multiple major versions deployed (v2/v3/v4), perform C1 reads on the NEWEST deployment separately — newer deployments often have weaker control surfaces than the legacy core.
- **C2.** Upgrade mechanism: transparent proxy / UUPS / Beacon / Diamond / immutable. Identify the proxy admin address. Check upgradeability of GOVERNANCE contracts too — a Governor / Aragon Voting / OZ Governor is often itself a proxy whose admin is the Timelock. Asymmetry: when fund-holding cores are immutable AND governance has no admin path that reaches them, an upgradable Governor/Timelock is T3-only and must NOT drag the verdict below green on that basis alone (see grade rules and the "immutable cores" caveat). Only call upgradability "mixed" if you can name a concrete function on the upgradable surface that reaches T1 or T2 on user funds.
- **C3.** EXECUTION PATH (enumerate every stage, in order, with delays in seconds). The operative path is usually a chain (voting → scheduler → timelock → executor; or governor → queue → execute; or Aragon Voting → DualGovernance → EmergencyProtectedTimelock → AdminExecutor). For each stage, record (a) the contract address, (b) the delay constant name + value in seconds, (c) the URL you read it from (block-explorer Read Contract OR `/api/contract/read?...&method=MIN_DELAY&block=<n>`). Do NOT stop at the first timelock-shaped contract — if its admin is itself called by another contract, keep walking. The grading delay is the SUM OF DELAYS ON THE UNCONTESTED FAST PATH (shortest time a proposal with no opposition can go from submission to executable). Dynamic / contested extensions (veto signaling, rage quit, escrow delay) are modifiers, not the basis — note them separately.
- **C4.** Enumerate EVERY multisig with reachable control — main proxy admin, emergency activation, emergency execution, reseal / pause, gate-seal committees, tiebreaker, per-module admins. For each Safe, fetch threshold + owners + version via `/api/safe/owners?chainId=<id>&address=0x...` (response includes raw eth_call data, so the URL is citable evidence). Enumerate ops/council/incentives multisigs even when off the upgrade path — record their scope so a reader can see they are NOT on the upgrade path. For each: (a) address, (b) threshold / total signers, (c) signer identities classified as insider (team, paid auditors under ongoing engagement, mandated service providers) vs non-insider (independent community members, unaffiliated researchers), (d) the specific power held (upgrade, pause, parameter, etc.).
- **C5.** On-chain governance: Governor / GovernorBravo / OZ Governor / Aragon Voting with token-weighted voting? Record proposal threshold, voting period, quorum, and the timelock delay between queue and execute. Every numeric constant must come from a Read Contract call you can link to, or be in unknowns[] with the C-code. If votingDelay / votingPeriod are denominated in BLOCKS, convert to seconds at the chain's CURRENT block time (Ethereum mainnet ≈ 12s post-Merge, not the 15s in older Compound/Bravo deployments) — cite both block count and converted seconds.
- **C6.** EMERGENCY POWERS: separate emergency-pause / guardian role with a different time cap or different actor than the main upgrade authority? Record it explicitly.
- **C7.** POWER TIER (blast radius). For each privileged path in C3–C6, classify the WORST thing that path can do, choosing the highest applicable tier. Cite the specific function name and any on-chain bound — tier claims without a named function are unsupported.
    - **T1 — FUND-CRITICAL**: replace implementation of contracts holding user funds; change AMM math / accounting / collateral logic; mint unbacked debt or shares; pause withdrawals; drain user-fund treasury; change oracle to attacker-controlled source; replace upgrade admin with EOA.
    - **T2 — ECONOMICALLY MATERIAL**: change fee parameters within bounded ranges; redirect protocol fees; add/remove markets / collateral types; bounded inflation or token mint within hard-capped schedule; spend protocol-owned (non-user) treasury.
    - **T3 — GOVERNANCE-INTERNAL**: change voting rules, quorum, voting period, proposal threshold; upgrade the Governor itself; rotate Timelock admin; mint governance tokens within a capped annual schedule.
    - **T4 — OPERATIONAL**: incentives distribution, grants, ENS / frontend canonicalization, deployment coordination, periphery router deprecation.
  The grade is set by the HIGHEST tier reachable on the uncontested fast path, not the median. State the tier and the binding function in the verdict.

### Read Contract discipline (applies to C3, C4, C5)

Every numeric constant cited (timelock delays, voting periods, multisig thresholds, quorum percentages) must come from EITHER (a) a block-explorer Read Contract URL, OR (b) a DeFiPunkd `/api/contract/read` or `/api/safe/owners` URL (preferred with `&block=<n>` for content-addressed evidence), OR (c) an unknowns[] entry with the C-code. Docs / blog posts are corroboration only — they cannot be the sole citation for a value that is also readable on-chain.

### Off-chain-only substitute hierarchy (when grading_basis="off-chain-only" — see preamble Rule 16)

When on-chain reads were genuinely unreachable this run, eligible off-chain substitutes in priority order:
1. Linked audit PDFs (admin roles, multisig members, timelock delays usually enumerated).
2. Governance forum posts that quote constants from a successful on-chain proposal (cite post URL + linked execution-tx URL).
3. Official protocol docs pages with named addresses and roles (must be on a domain owned by the protocol).
4. GitHub README / SECURITY.md / governance/*.md at a pinned commit SHA.

Forbidden substitutes: third-party blog posts, X / Twitter threads, search-result snippets, model memory. Required degradation: any C-code citing a numeric constant from docs/forum/audit prose ONLY must also carry an `unknowns[]` entry with `-offchain` suffix noting "value not re-read on-chain in this run; corroboration only".

### Grade rules (apply the timelock bar conditional on the highest C7 tier reachable on the fast path)

Security Council standard (used below): a multisig qualifies as "Security Council" only if ALL of: ≥7 signers, ≥51% threshold, ≥50% non-insider signers, every signer publicly announced. Failing any criterion = NOT a Security Council, regardless of signer reputation.

- **green**: highest reachable tier is T3 or T4 regardless of timelock; OR T2 reachable with uncontested-fast-path delay ≥7 days; OR T1 reachable only via immutable contracts (T1 is unreachable); OR T1 reachable with uncontested-fast-path delay ≥7 days combined with a Security Council multisig; OR T1 reachable with uncontested-fast-path delay ≥7 days through active on-chain governance with broad token distribution.
- **orange**: T2 reachable with uncontested-fast-path delay >0 but <7 days; OR T1 reachable with uncontested-fast-path delay >0 but <7 days; OR a multisig failing one or more Security Council criteria sits on a T1/T2 path; OR unclear upgrade authority on a T1/T2 path; OR governance with very short timelock or low quorum on a T1/T2 path.
- **red**: T1 reachable with no timelock by a single EOA or 2-of-3 multisig; OR a T1 upgrade admin that is not a smart contract you can audit.
- **unknown**: completed the checklist but still cannot determine the upgrade authority OR cannot classify the highest tier reachable on the main contracts.

Tiering caveats:
- "Bounded" must be enforced ON-CHAIN to count as T2. A function that sets fees with no upper-bound check is T1 — cite the bound check.
- Recurring T2 economic extraction (e.g. fee redirect with no rate limit) approaches T1 over time. A single proposal that can permanently redirect all future revenue is T1.
- T3 assumes the governance contract cannot itself authorize a T1/T2 action without going through the same timelock. If governance can self-upgrade to bypass the timelock, T3 collapses into T1.
- Do not downgrade tier by hand-waving ("realistically governance would never…"). Tier on what the contract permits, not what feels likely.

Notes:
- **Dynamic / dual-governance timelocks** (Lido, Compound escrow veto): the rubric grades on the uncontested path because that is the path most upgrades take. A dynamic extension that fires only under stake-weighted opposition is a real protection — name it in the green steel-man, but it does not lift an orange fast path into green; state the tension in the verdict.
- **Immutable cores with upgradable governance** (Uniswap-style): if fund-holding contracts are immutable and have no admin-reachable function moving / freezing / re-routing user funds, the highest reachable tier on the upgrade path is T3 — green regardless of timelock. Don't grade this orange just because the Governor is a proxy — that's a C2 fact, not a downgrade. Downgrade only applies if you can cite a concrete function on the upgradable surface that reaches T1 or T2 (privileged hook, upgradable factory controlling fund-routing, fee-switch redirecting protocol revenue without bound).
- **The 7-day bar** reflects the exit-window standard — users need notice after a queued upgrade to withdraw if they disagree. The ability-to-exit slice grades the exit side; this slice grades the delay side; both must hold for users to actually benefit from the delay.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "control",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Ability to exit unknown Unverified

Not yet assessed

Whether users can exit on their own terms if the team disappears or acts adversarially. Requires per-protocol review; not available at Phase 0.

No model has graded this dimension yet. Run the slice prompt through any LLM and submit the JSON — once ≥3 independent runs agree, the quorum bot merges the verdict here.

Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v1
- protocol.name:              Balancer V1
- protocol.chains:            Ethereum
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://docs.balancer.fi/reference/contracts/security.html#audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: ABILITY-TO-EXIT

Evaluate whether users can withdraw their funds on their own terms, even under adversarial admin conditions.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- E1. Enumerate every user-facing exit function in the main contracts: withdraw, redeem, burn, requestWithdrawal, claim, exit, etc. List them by name. Do NOT treat the contract as a monolith.
- E2. For EACH exit function in E1: identify its access modifiers and any pause guards (e.g. _checkResumed, whenNotPaused, onlyRole). Functions that gate REQUEST PLACEMENT often differ from functions that CLAIM FINALIZED FUNDS — check both separately.
- E3. For each pause guard: identify the role holder (which address holds PAUSE_ROLE / GUARDIAN / etc.) and the maximum pause duration. Specifically check whether PAUSE_INFINITELY (or equivalent uncapped pause) is callable, and which actor can call it (single multisig vs governance vote). For role-holder reads use https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=hasRole&args=0x...,0x... or &method=getRoleAdmin&args=0x.... For "is currently paused" checks use &method=paused or &method=isPaused&args=<resume-code>. Use the BARE method name (no parens). Cite the URL with &block=<n> in evidence[].
- E4. EMERGENCY vs GOVERNANCE pause distinction: many protocols have a fast-acting emergency pause capped at N days and a slow governance pause that can be indefinite. Record both paths separately if present, with their time caps and actor classes.
- E5. Queued redemption: documented maximum queue duration, daily withdrawal caps, whether the queue itself is pausable.
- E6. Forced-exit / escape-hatch / permissionless emergency-exit mechanism for adversarial-admin scenarios.
- E7. Frontend dependency: confirm exit functions are directly callable on-chain (e.g. via Etherscan write tab or a generic wallet) without the project's frontend.

Then write the steel-man section per Hard Rule 11. Common red-vs-orange tension on this slice: indefinite pause exists (suggests red) BUT the realistic emergency path is time-capped AND claims of already-finalized exits are not pause-gated (suggests orange). Resolve this by stating who can do what for how long, not by stopping at the worst-case sentence.

Grade rules:
- green   = permissionless exit; pause is either absent, narrowly scoped to clearly-described emergencies with auto-expiry, or capped at ≤7 days; no frontend dependency for exit; claims of already-finalized exits are not pause-gated under any path.
- orange  = pausable with broad scope OR indefinite pause is reachable only through governance vote (not unilateral admin action), OR queued redemption with documented max > 7 days, OR claims-of-finalized are exempt but new-request placement can be paused indefinitely by governance.
- red     = exit requires admin signature, OR ANY actor (including governance) can pause CLAIMS of finalized exits indefinitely, OR there is no on-chain exit function at all (purely custodial), OR pause is held by a single EOA / 2-of-3 multisig with no time cap.
- unknown = checklist incomplete after checking the sources above.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "ability-to-exit",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Autonomy unknown Unverified

No Phase-0 autonomy signal

No model has graded this dimension yet. Run the slice prompt through any LLM and submit the JSON — once ≥3 independent runs agree, the quorum bot merges the verdict here.

Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v1
- protocol.name:              Balancer V1
- protocol.chains:            Ethereum
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://docs.balancer.fi/reference/contracts/security.html#audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: AUTONOMY

Evaluate this protocol's autonomy: can a failure of anything outside its own contracts cause theft or loss of user principal, loss of unclaimed yield, or materially change the protocol's expected performance? "Autonomy" is not "has zero external touchpoints" — it is "external failures are either survivable or recoverable without user loss". This slice adapts DefiScan v1's Autonomy dimension; dependencies are one of several criteria, not the whole frame.

GUARDRAILS (read before grading):
- Category alone (Liquid Staking, Bridge, RWA Lending, Restaking, …) does NOT force a grade. A category is a hint about where to look; the grade must come from the concrete A1–A9 findings below.
- Base-chain consensus (Ethereum PoS, the chain's validator set, the canonical Deposit Contract at 0x00000000219ab540356cBB839Cbe05303d7705Fa) is the SUBSTRATE, not a dependency, for any protocol deployed on that chain. Do not list "depends on Ethereum" as a finding.
- Oracles or other integrations used by DOWNSTREAM protocols that happen to read this protocol's token (e.g. Chainlink stETH/USD consumed by Aave) are NOT this protocol's dependencies. Count only what THIS protocol's contracts call or trust on-chain. If a feed or contract appears in the protocol's docs only as reference material for third-party integrators, EXCLUDE IT ENTIRELY — do not log it as a finding even with a "peripheral" or "referenced only" caveat.
- "Upgradeable admin can change things" belongs to the CONTROL slice; only count it here when the upgrade surface lets governance silently swap an external dependency (see A9).
- Underlying-asset risk in opt-in, isolated markets is NOT autonomy-red on its own. When a protocol wraps third-party yield-bearing assets (LSTs, LRTs, ERC-4626 vaults, lending receipts, restaked tokens) into per-market silos that users explicitly choose, a failure of one underlying does NOT propagate to other markets and is risk the user opted into per-market. Record it under A4 with depth and propagation scope, but do not let it alone drive a red verdict — red requires an external dependency that cross-cuts the protocol or that the user did not opt into at deposit time. A failure mode that is "if the LRT you deposited is hacked, your principal in that LRT-backed market is impaired" is the underlying's autonomy story, not this protocol's; grade it on whether THIS protocol introduces additional dependencies on top.
- Sub-module enumeration is mandatory before grading. If the protocol ships distinct product lines or modules (e.g., a v2 core AMM plus a newer perps/funding-rate module, a lending pool plus a separate vault layer, an L1 core plus a cross-chain extension), enumerate each in findings and grade against the WORST module weighted by its share of TVS or its blast radius. A green core does not rescue an orange/red sub-module; conversely, a small red sub-module with capped TVS may bound the overall grade to orange. Name each module by its on-chain factory or router address. If you do not know whether a module exists, that is an unknowns[] entry, not silence. EXPLICIT TVS WEIGHTING: in the verdict, state each module's approximate share of total protocol TVS (use "~X%" if exact figures unavailable; check DeFiLlama or block-explorer balances on the module's main contract) and how that share informs the weighted grade. Format: "Module A holds ~X% of TVS (grade: <g>); Module B holds ~Y% (grade: <g>); weighted overall = <grade> because <reason>." If a red sub-module holds <5% of total TVS and is capped, the overall grade may be orange; if it holds >25%, the overall grade is red. Do not let qualitative reasoning substitute for the percentages — write the numbers.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. External contract calls. Enumerate every external contract the core contracts call or read from (oracles, price feeds, AMM pools, lending pools, staking/deposit contracts, yield wrappers). For each, identify the address, the provider, and what user-facing function of this protocol would break or mis-price if that external contract paused, mis-reported, or behaved adversarially. Grep for "oracle", "aggregator", "getPrice", "latestAnswer", "chainlink", "pyth", "redstone" as a starting point. To verify an oracle is live and what it currently reports, hit https://defipunkd.com/api/contract/read?chainId=<id>&address=<oracle>&method=latestAnswer (or &method=latestRoundData, &method=getPrice, &method=description; use the BARE method name without parens) and cite the URL — the response includes blockNumber/blockHash and rawReturnData, which is stronger evidence than a docs page about which feed "should" be used.
- A2. Off-chain actor committees reporting INTO the protocol. Oracle committees, guardian multisigs, DAO-selected validator sets acting as protocol reporters, exit-bus signers, fraud-proof challengers. For each, record committee size, quorum, who picks the members, and what mis-reporting could do (mint, burn, finalize withdrawal, freeze). Distinguish this from governance admins (control slice). NOTE ON STAKING PROTOCOLS: validator slashing and node-operator misbehavior are properties of the base-chain substrate, NOT external dependencies, when the operator set is diversified enough that a coordinated failure caps at <5% principal loss. Count validator/operator risk under A2 ONLY if the operator set is small, non-diversified, or lacks bonding / slashing-insurance / diversification mitigations. A curated set of 30+ independent operators with documented diversification falls under the mitigated path; a 3-operator LST does not. Do not cite the protocol's own risk disclosure as evidence that operator failure = principal loss unless you also check the diversification and bond mitigations.
- A3. Bridge / cross-chain messaging dependencies. Only count bridges that carry material TVL or are required for a core user flow. For each, name the bridge operator (canonical L1↔L2, LayerZero, Wormhole, Axelar, custom multisig), the trust model (canonical, optimistic, light-client, guardian set), and what fraction of TVL or users ride on it. "wstETH exists on 15 chains" is not a finding unless material TVL sits there. Before listing any non-primary chain deployment as a dependency, verify it is still operational as of analysis_date — retired or sunset deployments (e.g., Lido-on-Terra, Lido-on-Solana) belong in unknowns[] or should be omitted, never cited as a current dependency.
- A4. Nested collateral / restaking chains. For restaking / LRT / receipt-of-receipt designs: record the depth of the collateral chain, every actor with slashing or freezing power at each level, and whether a failure N levels deep propagates to user principal here.
- A5. Fork lineage (silent check). If DeFiLlama's forkedFrom is non-empty, record it as one finding and move on. If empty, do NOT add a placeholder finding; it adds noise.
- A6. Fallback mechanisms and circuit breakers. What catches an external failure? Sanity-check contracts on oracle reports, rebase bounds, pause paths triggered by bad prices, second-opinion oracles, max-per-block throughput caps, withdrawal queues that absorb bad reports. Record which A1–A4 risks are mitigated by which fallback, and which are unmitigated. For EACH fallback you cite, state its activation status explicitly: (i) LIVE and enforcing on-chain today, (ii) DEPLOYED but not yet wired / activated (e.g., interface exists but the address is zero or the role is unassigned), or (iii) DOCUMENTED / PROPOSED only (forum post, LIP draft, audit pending). Only (i) counts as mitigation for the grade. A fallback in state (ii) or (iii) should be noted but must not reduce the risk in your steel-man or verdict. If you cannot determine activation status, add an unknowns[] entry rather than assume it is live.
- A7. Sequencer / L1-liveness dependency BEYOND the base-chain substrate. SCOPE — sequencer risk only counts here when the protocol IS its own L2/L3 appchain or app-rollup, where the sequencer is part of the protocol's own stack and a freeze is a protocol-level outage. A protocol permissionlessly deployed on a third-party L2 inherits that L2's sequencer as substrate, not an A7 dependency. Record the sequencer/DA trust model when A7 applies.
- A8. Keeper / relayer / off-chain bot liveness. Protocols that need permissionless-but-necessary off-chain actors (liquidation bots, auto-compounders, deposit relayers, intent solvers). Record whether the role is permissionless, what degrades if nobody runs it (yield paused, bad debt accumulates, positions go stale), and whether the failure mode is graceful or catastrophic.
- A9. Governance-mutable dependency surface. Can an admin or DAO action silently INTRODUCE a NEW EXTERNAL dependency — swap the oracle address to a different provider, register a new staking module that calls an untrusted contract, add a new bridge, route SY through a new external vault — without an exit window for users? Check the upgrade / router / module-registry contracts. Answer: which EXTERNAL dependencies are governance-mutable, who holds that power, and whether there is a timelock or exit window. SCOPE LIMIT — read carefully: A9 is about the *external dependency surface*, not the upgrade surface in general. "The proxyAdmin / EOA can upgrade the router implementation to arbitrary bytecode" is a CONTROL-slice finding (admin can rug), NOT an autonomy-A9 finding. A9 fires only when the upgrade specifically swaps out or adds a contract that THIS protocol calls or trusts (e.g., changing the oracle address from Chainlink to a malicious feed, registering a new SY adapter that points to a third-party vault, redirecting a bridge endpoint). If the only finding is "admin can change implementation," do not log it under A9 and do not let it drive the autonomy grade — note it under control instead. The autonomy-relevant version of the same upgrade key is "admin can swap [specific external address X] without timelock"; that requires identifying the specific external dependency that becomes mutable.

STEEL-MAN (per Hard Rule 13): write one-sentence strongest arguments for red, orange, and green using the A1–A9 findings.

IMPACTED TVS ESTIMATE: the headline MUST include a rough impacted-TVS figure — the fraction of protocol TVS that could be lost or frozen if the worst-unmitigated dependency you identified failed. Use "~X%" if exact numbers are unavailable, "<1%" for de minimis, "unclear" only if A1–A9 left the question genuinely open (in which case grade=unknown is usually correct). Do NOT substitute qualitative phrases like "significant" — give a number or bracket.

GRADE ANCHORS (mapped to DefiScan v1 stages):
- green = Stage 2 equivalent. Failure of any external dependency cannot cause loss of user principal or unclaimed yield. Either there are no material external dependencies, or every critical one has a documented fallback (A6) that keeps users whole. Governance cannot silently introduce new dependencies without an exit window (A9). Impacted TVS under any single-dependency failure: effectively 0.
- orange = Stage 1 equivalent. Failure of some external dependency can cause loss of unclaimed yield, or can materially change expected performance (pause withdrawals, freeze positions, degrade price quality), but cannot cause loss of principal. Committee-based oracles with sanity checks, canonical-only bridges, fallback paths that exist but are incomplete, or governance-mutable dependencies protected by a ≥7-day timelock. Impacted TVS is bounded and recoverable.
- red = Stage 0 equivalent. Failure of an external dependency CAN cause theft or loss of principal. Examples: single-provider oracle with no sanity check or fallback, material TVL on a non-canonical bridge with a guardian multisig, governance can hot-swap oracles or add staking modules with no timelock or exit window, unmitigated keeper-liveness dependency where positions become insolvent if bots stop. Impacted TVS is material.
- unknown = checklist incomplete after inspecting source + verified contracts. Prefer unknown over guessing when A1/A6 could not be reconstructed. RESERVE unknown for cases where the CORE ARCHITECTURE itself is unverifiable — not for cases where you merely cannot enumerate every per-market dependency in a multi-market protocol. If the core router/factory/oracle architecture is verifiable on-chain and you can determine whether the core requires external dependencies, grade the architecture even when an exhaustive per-market external-dependency census is infeasible. Acknowledge the per-market gap in unknowns[] but still issue a grade. Refusing to grade a multi-market protocol because you cannot list every SY/vault/market is over-use of unknown; grade the architecture and say so.

PROMPT-META CHECK (per Hard Rule 17): before finalizing, verify the verdict cites concrete contract addresses, docs, or code — not the rubric itself. If your verdict says "the protocol belongs to a category the rubric marks red", rewrite it with the A1–A9 finding that actually justifies the grade, or drop to grade=unknown.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "autonomy",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Open Access unknown Unverified

Not yet assessed

No model has graded this dimension yet. Run the slice prompt through any LLM and submit the JSON — once ≥3 independent runs agree, the quorum bot merges the verdict here.

Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-v1
- protocol.name:              Balancer V1
- protocol.chains:            Ethereum
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi/
- protocol.github:            balancer
- protocol.audit_links:       https://docs.balancer.fi/reference/contracts/security.html#audits, https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: OPEN-ACCESS

Evaluate who is allowed to use the protocol and whether any of that permission is granted off-chain.

Scope: this slice is about ADMISSION — who can enter, exit, or transact. Operator LIVENESS (what breaks if keepers/oracles go offline) is assessed in the dependencies slice and is out of scope for the grade here. You may note operator dependencies as context, but do not let "the protocol halts if operator X disappears" drive the access grade on its own; that belongs in dependencies. Source verification / contract verification on block explorers is assessed in the verifiability slice and is out of scope here — do NOT let "contract is unverified" drive the access grade.

Framing: the smart contracts are the access layer; frontends are UX. A permissionless contract is reachable by any client (SDK, third-party UI, aggregator, wallet integration). Frontend ToS, IP geo-blocking, and wallet screening are publisher policies on one specific client — they are reported as context but do NOT determine the grade. The grade hinges on (1) what the contract itself permits, and (2) whether the protocol is practically reachable without the official publisher's cooperation.

Meta-check before finalizing: if your verdict cites phrases from this prompt as evidence ("the protocol meets the 'credible alternatives' condition", "this fits the 'documented fallback' rule"), redo the verdict. The prompt describes the rubric; evidence must come from the protocol. A verdict should cite what the protocol does, not what the rubric says.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. Whitelist / allowlist modifiers in user-facing entry points. Grep for "onlyWhitelisted", "onlyRole", "allowlist", "isAccredited", "isKYCed". Note which functions are gated and who can add/remove from the list.
- A2. Off-chain operators in the admission path: keepers, sequencers, privileged relayers, oracle posters whose approval is required to admit a user action (not just to keep the protocol live). For each, identify whether the role is held by a single operator, a permissioned committee, or is permissionless. Enumerate per user-facing function class (deposit vs withdraw-request vs claim-finalized vs transfer) which ones require operator approval to be admitted, and which ones admit users unconditionally. A function whose placement is unconditional but whose downstream settlement depends on an operator is an admission-permissionless function — flag the liveness dependency as context and defer its grading weight to the dependencies slice.
- A3. Frontend restrictions on the official interface — record as context, not as a grade lever. Distinguish:
    - A3-passive: boilerplate ToS clauses (sanctions attestation, restricted-territory self-certification, VPN-circumvention prohibition, "comply with applicable law" eligibility, age of majority).
    - A3-active: runtime enforcement — IP-based geo-blocking, wallet-address screening against a sanctions oracle (Chainalysis, TRM, Elliptic), KYC wall, rendering-blocking jurisdiction banner.
  Record findings under the correct tier. Quote ToS text or banner text in evidence[].shows. These findings populate the headline and rationale but do NOT move the grade by themselves; the grade is set by A1, A2, and the A3b path check below.
- A3b. Independent access paths (the operative grade input). Enumerate paths that do not require the official publisher's cooperation:
    - Published SDK / library / CLI for direct contract interaction.
    - Third-party frontends operated by separate legal entities.
    - Wallet-integrated access (MetaMask Swaps, Safe apps, etc.).
    - DEX / lending / yield aggregators that route through the contracts.
  Record at least one concrete link per path that exists. The protocol does NOT have to self-document these — the test is existence, not UX cost. An A3b-i redistribution of the official UI bound by the same ToS does NOT count as an independent path.
- A4. Sanctions / compliance tooling at the contract level: does the protocol check addresses against OFAC lists or similar on-chain blocklists in the contract itself? (Frontend-only screening belongs in A3.)
- A5. Differentiate read access vs write access: many protocols are read-permissionless (anyone can view state) but write-gated (only certain addresses can deposit/borrow). Record both.
- A6. ToS / Legal links: locate them on the website and produce a VERBATIM quote of any jurisdictional, sanctions, or eligibility clause in evidence[].shows. If you cannot extract the clause text verbatim (SPA render failure, paywall, dead link, etc.), do NOT paraphrase or infer from general knowledge — record the ToS URL in unknowns[] with the reason extraction failed. Assertions about ToS content without a verbatim quote will be downweighted by reviewers.

Then write the steel-man section per Hard Rule 11.

Grade rules (admission-focused; liveness concerns belong in dependencies; source verification belongs in verifiability):
- green   = no contract-level whitelist/KYC on user entry/exit; no operator approval required to admit a user action; AND at least one independent A3b path exists (published SDK, third-party frontend, wallet integration, or aggregator routing). Frontend ToS posture and A3-active enforcement on the official UI do NOT block green when contracts are permissionless and an independent path exists — they are reported as context.
- orange  = contracts admit users unconditionally, BUT the protocol is operationally captured by the official publisher: no published SDK, no third-party frontend, no wallet integration, no aggregator routing. The contract is theoretically open but practically reachable only through the official UI. Also applies when admission requires approval from a permissioned committee that is governance-managed with a documented replacement procedure.
- red     = contract-level whitelist / KYC on user entry/exit, OR admission of a core user action requires approval from a single privileged operator or a small committee with no documented replacement procedure, OR enforces an on-chain blocklist updatable by a single party.
- unknown = checklist incomplete after checking the sources above.

Default-grade guidance: when contracts are fully permissionless AND any A3b independent path exists, the default grade is green regardless of frontend ToS or A3-active enforcement on the official UI. Frontend geo-blocking, sanctions-oracle wallet screening, and ToS sanctions clauses are publisher policies on one client and are reported in findings/headline as context, not as grade levers. To grade orange on operational-capture grounds, the auditor must affirmatively show that ALL independent paths are absent or also gated.

Guideline on committees: where admission depends on a multi-operator committee, the relevant axes are (a) set size, (b) whether replacement/rotation is governed on-chain, (c) whether the replacement procedure is publicly documented. A large set with on-chain governance replacement should not be graded as a single-party operator even if rotation is not instantaneous. A small set with informal replacement should be treated as a single-party operator.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "open-access",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Stage

Preview of the Phase-3 maturity framework. DeFiPunk'd will adopt DeFiScan v2's stages verbatim; the section is rendered below in its intended shape so the structure is visible today.

Balancer V1 has not yet been assessed under the DeFiScan v2 stage framework.

The walkaway test is the central criterion. Once stages land, protocols reach Stage 1 only if users can exit in the presence of malicious operators even when the emergency council disappears.

Scope of assessment

Stage 0 requirements pending

Stage 1 requirements pending

Stage 2 requirements pending

Learn more about DeFiScan v2 stages →

TVL $999.3K

Type Dexs

Chains xDai, Ethereum, Arbitrum, Base

View on DeFiLlama ↗

Control Exit Autonomy Open Access Verifiable

Verifiability tentative Open source + 3 audits
Control unknown Not yet assessed
Ability to exit unknown Not yet assessed
Autonomy unknown No Phase-0 autonomy signal
Open Access unknown Not yet assessed

Control criteria

Upgradeability — Bug bounty immunefi.com Governance forum — Docs —

About

Balancer CoW AMM is a dexs protocol on xDai.

Risk analysis

One card per dimension, sorted by severity. Only Verifiability and Autonomy carry automated signals in Phase 0. See methodology for scope.

Audit a dimension yourself · DEFI@home Contribute an LLM-run assessment — any model, any dimension. Three agreeing runs merge automatically into the public record.

Address discovery 0 addresses on file · 0 runs ⚑ Run first Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-cow-amm
- protocol.name:              Balancer CoW AMM
- protocol.chains:            xDai, Ethereum, Arbitrum, Base
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: DISCOVERY

You are a cataloguer, not a judge. Your job is to surface every contract address that could plausibly belong to this protocol's control or fund-holding surface, each backed by a citation. `grade` is ALWAYS `"unknown"` for discovery submissions — there is no green/orange/red rubric here. The five evaluation slices that run after you (control, ability-to-exit, autonomy, open-access, verifiability) consume your output via the addressBook ratchet — every address you record becomes a pre-built surfacer URL on the next run; every address you miss costs them a tool call.

Width beats depth. A `role: "other"` entry with one cited URL beats omitting it. Downstream slices will discard out-of-scope entries; they cannot rediscover what you fail to enumerate without paying the same cost again.

(Step 0 capability probe lives in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every D-code below must appear in evidence[] OR unknowns[])

- **D1. Block-explorer name-tag search per chain.** For each chain in `protocol.chains`, search the canonical block explorer for the protocol's name tag — `https://etherscan.io/searchHandler?term=<query>` and the per-chain explorers (basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network). When direct fetch is blocked, use `site:<explorer> <protocol_name>` via search grounding. Record every address that surfaces with the protocol's tag, plus neighbouring "Token Contract" / "Multisig" / "Timelock" labels.

- **D2. Official deployments doc.** From `protocol.website` and the docs site, locate the canonical "Deployed contracts" / "Addresses" / "Contracts" / "Deployments" page (often at `/docs/deployments`, `/docs/addresses`, `/dashboard/contracts`). Cite the URL, record every address listed with its named role, and set `protocol_metadata.deployed_contracts_doc` to this URL.

- **D3. Audit PDFs.** From `protocol.audit_links` (and any audits surfaced by D2), open each. Most reports include a "Scope" / "Contracts in scope" address table in the first 5 pages. Extract every in-scope address with its labelled role. If the audit predates the current deployment, record the addresses anyway with role suffixed `(audit-era)` so downstream slices know to re-verify.

- **D4. GitHub deployment artifacts.** From `protocol.github`, walk the repo at a pinned commit SHA looking for: Foundry `broadcast/<script>/<chainId>/run-latest.json` (`transactions[].contractAddress` per chainId); hardhat-deploy `deployments/<network>/<Contract>.json` (`address` field); manual indexes (`deployments.json`, `addresses.json`, `contracts.json`, `networks.json`); markdown indexes (`docs/deployments.md`, `README.md` tables). Cite the file URL with the commit SHA; pin SHAs (`?ref=<sha>`) so the citation is content-addressed.

- **D5. Multi-chain enumeration.** If `protocol.chains.length > 1`, repeat D1–D4 per chain. Cross-chain deployments of the same logical contract get SEPARATE `admin_addresses[]` entries — one per chain. The chain field is part of the identity; do not collapse. If a chain has zero results, record `"D5: chain <name>: zero addresses surfaced from <sources tried>"` in unknowns[].

- **D6. Factory-discovered children.** For factory addresses surfaced in D1–D4, fetch the enumeration view via the read API (`/api/contract/read?...&method=allPools` / `getPool` / `getMarket` / `getVault`) and record each child with role like `"pool (from factory <0xFactory>)"`. **Cap at 50 children per factory.** Protocols with thousands of pools (Uniswap, Sushi) need dedicated ingestion — record the factory + the cap notice in unknowns[].

- **D7. Role taxonomy.** Every `admin_addresses[]` entry's `role` uses this controlled vocabulary (free-text suffixes OK for disambiguation, e.g. `"multisig (treasury)"`, but the leading token must match):

    `owner | admin | proxy_admin | governor | timelock | guardian | multisig | treasury | oracle | factory | router | token | pool | vault | other`

  Tentative classifications are encouraged. `actor_class` ∈ `eoa | multisig | timelock | governance | unknown` — use `unknown` when you found the address but didn't read its bytecode.

- **D8. Ratchet output integrity.** Every address in `admin_addresses[]` must trace to ≥1 fetched URL in evidence[]. Snippet-only sightings go in unknowns[] with a `D8` code, NOT in admin_addresses[].

### Discovery rationale framing

- `rationale.findings`: one entry per D-code, terse, factual. Per-address detail belongs in evidence[] and admin_addresses[], not here. Example: `"D1: 8 addresses surfaced from etherscan.io name-tag search for 'Aave V3'"`.
- `rationale.steelman`: ALWAYS null.
- `rationale.verdict`: one short line summarizing what corpora were walked and how many addresses were catalogued.
- `headline`: factual and quantitative — `"24 contracts catalogued across Ethereum, Arbitrum, and Base; 6 governance/admin and 18 protocol contracts."`.
- `short_headline`: under 60 chars — `"24 contracts across 3 chains"`.

### What discovery is NOT

- Not a verdict slice. `grade` must be `"unknown"`.
- Not exhaustive enumeration of leaf assets — record the factory + cap and move on (see D6).
- Not classification of trust assumptions — whether a multisig threshold is safe / timelock delay is sufficient / proxy admin is an EOA is the control slice's job.
- Not address-book reconciliation: when addressBook is non-empty, EXTEND it (find addresses prior runs missed) rather than re-cite the same addresses; re-cite only when you have new evidence for a refined role.

### protocol_metadata side-effects

While walking the corpora, populate every `protocol_metadata` field you can support with citations: `github`, `docs_url`, `audits` (one per D3 audit walked), `governance_forum`, `bug_bounty_url`, `security_contact`, `deployed_contracts_doc` (URL from D2), `upgradeability` (best-effort), and `about` (2–4 sentences sourced from docs/website, not memory). Discovery is the natural home for these — evaluation slices should not have to rediscover them.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "discovery",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Verifiability Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-cow-amm
- protocol.name:              Balancer CoW AMM
- protocol.chains:            xDai, Ethereum, Arbitrum, Base
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: VERIFIABILITY

Evaluate whether an outsider can independently confirm what the deployed code does.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- V1. For each address you assess: is the bytecode verified on the chain's block explorer? Record the "Contract Source Code Verified" indicator. https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x... returns this as a top-level "verified" boolean plus "abiSource" ("etherscan" / "sourcify") and an inline ABI — useful when the explorer page is rate-limited. If the contract is a proxy, verify BOTH the proxy contract AND the current implementation contract. The same /api/contract/abi response auto-resolves proxies and includes a "proxy.implementation" address when present, so one call covers V1 + V6 in one shot. An explorer "Similar Match" on a well-known proxy pattern (Aragon AppProxyUpgradeable, ERC1967Proxy, OssifiableProxy, OZ TransparentUpgradeableProxy) is expected for that pattern and does NOT count as a verification gap on its own — what matters is that the implementation is independently verified.
- V2. Source-to-repo correspondence: for each verified contract, attempt to find a matching commit in the linked GitHub repos. Record evidence[].commit on a match. Independent compile/bytecode-match is NOT required for green — a recognized public repo whose structure and file contents correspond to the explorer-visible source is sufficient. If you did not pin a commit SHA or run a bytecode diff, record that plainly in unknowns[] and proceed; it is a scope limit, not a downgrade signal.
- V3. Audit coverage: for each URL in protocol.audit_links, open it and record: auditor name, audit date, the specific contracts / commit in scope. Flag audits that predate the current deployment by >6 months without a follow-up review.
- V4. Auditor recognition: the following firms are broadly recognized in Solidity: Trail of Bits, Zellic, Spearbit, OpenZeppelin, ConsenSys Diligence, Certora (formal verification), Quantstamp, Halborn, Peckshield, Sigma Prime, ChainSecurity, Ackee Blockchain, MixBytes, Statemind. Unknown firms are orange-at-best for any green-grade claim. Name the firm explicitly in evidence[].
- V5. Post-audit drift: compare the most recent in-scope audit(s) against the currently-deployed source, weighted by what each contract does and by what the changes actually contain. SCOPE — drift only downgrades the grade when ALL of the following hold: (i) the drifting contract is fund-custody / settlement / accounting-critical (NOT a peripheral router, lens, quoter, or pure-view contract that holds no balances); (ii) the changes are material — new functions, modified access control, modified accounting, modified fund flow — and not refactors / struct relocations / import reorgs / build or CI fixes / formatting; (iii) no later audit, fix-audit, or differential audit from a recognized firm covers the changed files (audits often pin a pre-fix commit while a follow-up reviews the delta — match by file scope, not by commit-hash equality). When you cite drift as a downgrade reason, name the specific behavior change (function added, role granted, accounting formula altered) — "N commits ahead" or "+X/-Y LOC" alone is not evidence of material drift. If you have not sampled the diff content (e.g. via the GitHub compare view or the top commits in the window), record drift as an unknowns[] entry rather than auto-downgrading; commit-count and LOC are starting signals, not findings.
- V6. Implementation vs proxy: a verified proxy with an unverified implementation is effectively unverified. State whether the implementation is verified separately.

EVIDENCE DISCIPLINE (read before writing findings[]):
- Do not assert a specific deploy-commit SHA, bytecode equivalence, or "identical to audited commit" unless you actually fetched the artifact that shows it (e.g., a deployed-addresses JSON you opened, an explorer page you read). Inferred or plausible matches belong in unknowns[], never in findings[] or evidence[].
- Evidence[] entries must correspond to pages/files you actually retrieved this run. A URL you did not open is not evidence.

Then write the steel-man section per Hard Rule 11.

Grade rules:
- green   = deployed bytecode verified on the explorer (proxy AND implementation if proxied; "Similar Match" on a standard proxy pattern is fine per V1), a public source repo exists whose contents correspond to the explorer-visible source, AND ≥1 audit from a recognized firm covering the currently-deployed contracts (≤6 months of drift OR drift was re-audited). A missing local compile-match is not a downgrade — record it in unknowns[] and still grade green if the other conditions hold.
- orange  = verified but with visible drift from the public repo, OR audit scope is stale relative to deployment, OR only minor / unknown-firm audits exist, OR only some of the main contracts are verified, OR proxy verified but implementation only partially verified.
- red     = unverified bytecode (or verified proxy with unverified implementation), OR no audit in protocol.audit_links, OR no public repo.
- unknown = reserved for when the protocol's verifiability posture genuinely cannot be assessed (e.g., explorer and repo both inaccessible for this protocol). Do NOT use unknown merely because you, the analyst, could not run a particular check such as a bytecode diff — that goes in unknowns[] while the grade is still assigned from the evidence you do have.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "verifiability",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Control Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-cow-amm
- protocol.name:              Balancer CoW AMM
- protocol.chains:            xDai, Ethereum, Arbitrum, Base
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: CONTROL

Evaluate who can change the protocol's rules, how fast, and how broadly.

(Step 0 capability probe and the off-chain-only fallback live in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[])

- **C1.** For each address you assess: who is the contract owner / admin / pendingAdmin / governor — read these via the block explorer's "Read Contract" tab OR `https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=owner` (BARE method names: `&method=owner`, `&method=admin`, `&method=pendingOwner`, `&method=governor`). For Safes use `/api/safe/owners?chainId=<id>&address=0x...`. When a protocol has multiple major versions deployed (v2/v3/v4), perform C1 reads on the NEWEST deployment separately — newer deployments often have weaker control surfaces than the legacy core.
- **C2.** Upgrade mechanism: transparent proxy / UUPS / Beacon / Diamond / immutable. Identify the proxy admin address. Check upgradeability of GOVERNANCE contracts too — a Governor / Aragon Voting / OZ Governor is often itself a proxy whose admin is the Timelock. Asymmetry: when fund-holding cores are immutable AND governance has no admin path that reaches them, an upgradable Governor/Timelock is T3-only and must NOT drag the verdict below green on that basis alone (see grade rules and the "immutable cores" caveat). Only call upgradability "mixed" if you can name a concrete function on the upgradable surface that reaches T1 or T2 on user funds.
- **C3.** EXECUTION PATH (enumerate every stage, in order, with delays in seconds). The operative path is usually a chain (voting → scheduler → timelock → executor; or governor → queue → execute; or Aragon Voting → DualGovernance → EmergencyProtectedTimelock → AdminExecutor). For each stage, record (a) the contract address, (b) the delay constant name + value in seconds, (c) the URL you read it from (block-explorer Read Contract OR `/api/contract/read?...&method=MIN_DELAY&block=<n>`). Do NOT stop at the first timelock-shaped contract — if its admin is itself called by another contract, keep walking. The grading delay is the SUM OF DELAYS ON THE UNCONTESTED FAST PATH (shortest time a proposal with no opposition can go from submission to executable). Dynamic / contested extensions (veto signaling, rage quit, escrow delay) are modifiers, not the basis — note them separately.
- **C4.** Enumerate EVERY multisig with reachable control — main proxy admin, emergency activation, emergency execution, reseal / pause, gate-seal committees, tiebreaker, per-module admins. For each Safe, fetch threshold + owners + version via `/api/safe/owners?chainId=<id>&address=0x...` (response includes raw eth_call data, so the URL is citable evidence). Enumerate ops/council/incentives multisigs even when off the upgrade path — record their scope so a reader can see they are NOT on the upgrade path. For each: (a) address, (b) threshold / total signers, (c) signer identities classified as insider (team, paid auditors under ongoing engagement, mandated service providers) vs non-insider (independent community members, unaffiliated researchers), (d) the specific power held (upgrade, pause, parameter, etc.).
- **C5.** On-chain governance: Governor / GovernorBravo / OZ Governor / Aragon Voting with token-weighted voting? Record proposal threshold, voting period, quorum, and the timelock delay between queue and execute. Every numeric constant must come from a Read Contract call you can link to, or be in unknowns[] with the C-code. If votingDelay / votingPeriod are denominated in BLOCKS, convert to seconds at the chain's CURRENT block time (Ethereum mainnet ≈ 12s post-Merge, not the 15s in older Compound/Bravo deployments) — cite both block count and converted seconds.
- **C6.** EMERGENCY POWERS: separate emergency-pause / guardian role with a different time cap or different actor than the main upgrade authority? Record it explicitly.
- **C7.** POWER TIER (blast radius). For each privileged path in C3–C6, classify the WORST thing that path can do, choosing the highest applicable tier. Cite the specific function name and any on-chain bound — tier claims without a named function are unsupported.
    - **T1 — FUND-CRITICAL**: replace implementation of contracts holding user funds; change AMM math / accounting / collateral logic; mint unbacked debt or shares; pause withdrawals; drain user-fund treasury; change oracle to attacker-controlled source; replace upgrade admin with EOA.
    - **T2 — ECONOMICALLY MATERIAL**: change fee parameters within bounded ranges; redirect protocol fees; add/remove markets / collateral types; bounded inflation or token mint within hard-capped schedule; spend protocol-owned (non-user) treasury.
    - **T3 — GOVERNANCE-INTERNAL**: change voting rules, quorum, voting period, proposal threshold; upgrade the Governor itself; rotate Timelock admin; mint governance tokens within a capped annual schedule.
    - **T4 — OPERATIONAL**: incentives distribution, grants, ENS / frontend canonicalization, deployment coordination, periphery router deprecation.
  The grade is set by the HIGHEST tier reachable on the uncontested fast path, not the median. State the tier and the binding function in the verdict.

### Read Contract discipline (applies to C3, C4, C5)

Every numeric constant cited (timelock delays, voting periods, multisig thresholds, quorum percentages) must come from EITHER (a) a block-explorer Read Contract URL, OR (b) a DeFiPunkd `/api/contract/read` or `/api/safe/owners` URL (preferred with `&block=<n>` for content-addressed evidence), OR (c) an unknowns[] entry with the C-code. Docs / blog posts are corroboration only — they cannot be the sole citation for a value that is also readable on-chain.

### Off-chain-only substitute hierarchy (when grading_basis="off-chain-only" — see preamble Rule 16)

When on-chain reads were genuinely unreachable this run, eligible off-chain substitutes in priority order:
1. Linked audit PDFs (admin roles, multisig members, timelock delays usually enumerated).
2. Governance forum posts that quote constants from a successful on-chain proposal (cite post URL + linked execution-tx URL).
3. Official protocol docs pages with named addresses and roles (must be on a domain owned by the protocol).
4. GitHub README / SECURITY.md / governance/*.md at a pinned commit SHA.

Forbidden substitutes: third-party blog posts, X / Twitter threads, search-result snippets, model memory. Required degradation: any C-code citing a numeric constant from docs/forum/audit prose ONLY must also carry an `unknowns[]` entry with `-offchain` suffix noting "value not re-read on-chain in this run; corroboration only".

### Grade rules (apply the timelock bar conditional on the highest C7 tier reachable on the fast path)

Security Council standard (used below): a multisig qualifies as "Security Council" only if ALL of: ≥7 signers, ≥51% threshold, ≥50% non-insider signers, every signer publicly announced. Failing any criterion = NOT a Security Council, regardless of signer reputation.

- **green**: highest reachable tier is T3 or T4 regardless of timelock; OR T2 reachable with uncontested-fast-path delay ≥7 days; OR T1 reachable only via immutable contracts (T1 is unreachable); OR T1 reachable with uncontested-fast-path delay ≥7 days combined with a Security Council multisig; OR T1 reachable with uncontested-fast-path delay ≥7 days through active on-chain governance with broad token distribution.
- **orange**: T2 reachable with uncontested-fast-path delay >0 but <7 days; OR T1 reachable with uncontested-fast-path delay >0 but <7 days; OR a multisig failing one or more Security Council criteria sits on a T1/T2 path; OR unclear upgrade authority on a T1/T2 path; OR governance with very short timelock or low quorum on a T1/T2 path.
- **red**: T1 reachable with no timelock by a single EOA or 2-of-3 multisig; OR a T1 upgrade admin that is not a smart contract you can audit.
- **unknown**: completed the checklist but still cannot determine the upgrade authority OR cannot classify the highest tier reachable on the main contracts.

Tiering caveats:
- "Bounded" must be enforced ON-CHAIN to count as T2. A function that sets fees with no upper-bound check is T1 — cite the bound check.
- Recurring T2 economic extraction (e.g. fee redirect with no rate limit) approaches T1 over time. A single proposal that can permanently redirect all future revenue is T1.
- T3 assumes the governance contract cannot itself authorize a T1/T2 action without going through the same timelock. If governance can self-upgrade to bypass the timelock, T3 collapses into T1.
- Do not downgrade tier by hand-waving ("realistically governance would never…"). Tier on what the contract permits, not what feels likely.

Notes:
- **Dynamic / dual-governance timelocks** (Lido, Compound escrow veto): the rubric grades on the uncontested path because that is the path most upgrades take. A dynamic extension that fires only under stake-weighted opposition is a real protection — name it in the green steel-man, but it does not lift an orange fast path into green; state the tension in the verdict.
- **Immutable cores with upgradable governance** (Uniswap-style): if fund-holding contracts are immutable and have no admin-reachable function moving / freezing / re-routing user funds, the highest reachable tier on the upgrade path is T3 — green regardless of timelock. Don't grade this orange just because the Governor is a proxy — that's a C2 fact, not a downgrade. Downgrade only applies if you can cite a concrete function on the upgradable surface that reaches T1 or T2 (privileged hook, upgradable factory controlling fund-routing, fee-switch redirecting protocol revenue without bound).
- **The 7-day bar** reflects the exit-window standard — users need notice after a queued upgrade to withdraw if they disagree. The ability-to-exit slice grades the exit side; this slice grades the delay side; both must hold for users to actually benefit from the delay.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "control",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Ability to exit Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-cow-amm
- protocol.name:              Balancer CoW AMM
- protocol.chains:            xDai, Ethereum, Arbitrum, Base
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: ABILITY-TO-EXIT

Evaluate whether users can withdraw their funds on their own terms, even under adversarial admin conditions.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- E1. Enumerate every user-facing exit function in the main contracts: withdraw, redeem, burn, requestWithdrawal, claim, exit, etc. List them by name. Do NOT treat the contract as a monolith.
- E2. For EACH exit function in E1: identify its access modifiers and any pause guards (e.g. _checkResumed, whenNotPaused, onlyRole). Functions that gate REQUEST PLACEMENT often differ from functions that CLAIM FINALIZED FUNDS — check both separately.
- E3. For each pause guard: identify the role holder (which address holds PAUSE_ROLE / GUARDIAN / etc.) and the maximum pause duration. Specifically check whether PAUSE_INFINITELY (or equivalent uncapped pause) is callable, and which actor can call it (single multisig vs governance vote). For role-holder reads use https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=hasRole&args=0x...,0x... or &method=getRoleAdmin&args=0x.... For "is currently paused" checks use &method=paused or &method=isPaused&args=<resume-code>. Use the BARE method name (no parens). Cite the URL with &block=<n> in evidence[].
- E4. EMERGENCY vs GOVERNANCE pause distinction: many protocols have a fast-acting emergency pause capped at N days and a slow governance pause that can be indefinite. Record both paths separately if present, with their time caps and actor classes.
- E5. Queued redemption: documented maximum queue duration, daily withdrawal caps, whether the queue itself is pausable.
- E6. Forced-exit / escape-hatch / permissionless emergency-exit mechanism for adversarial-admin scenarios.
- E7. Frontend dependency: confirm exit functions are directly callable on-chain (e.g. via Etherscan write tab or a generic wallet) without the project's frontend.

Then write the steel-man section per Hard Rule 11. Common red-vs-orange tension on this slice: indefinite pause exists (suggests red) BUT the realistic emergency path is time-capped AND claims of already-finalized exits are not pause-gated (suggests orange). Resolve this by stating who can do what for how long, not by stopping at the worst-case sentence.

Grade rules:
- green   = permissionless exit; pause is either absent, narrowly scoped to clearly-described emergencies with auto-expiry, or capped at ≤7 days; no frontend dependency for exit; claims of already-finalized exits are not pause-gated under any path.
- orange  = pausable with broad scope OR indefinite pause is reachable only through governance vote (not unilateral admin action), OR queued redemption with documented max > 7 days, OR claims-of-finalized are exempt but new-request placement can be paused indefinitely by governance.
- red     = exit requires admin signature, OR ANY actor (including governance) can pause CLAIMS of finalized exits indefinitely, OR there is no on-chain exit function at all (purely custodial), OR pause is held by a single EOA / 2-of-3 multisig with no time cap.
- unknown = checklist incomplete after checking the sources above.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "ability-to-exit",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Autonomy Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-cow-amm
- protocol.name:              Balancer CoW AMM
- protocol.chains:            xDai, Ethereum, Arbitrum, Base
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: AUTONOMY

Evaluate this protocol's autonomy: can a failure of anything outside its own contracts cause theft or loss of user principal, loss of unclaimed yield, or materially change the protocol's expected performance? "Autonomy" is not "has zero external touchpoints" — it is "external failures are either survivable or recoverable without user loss". This slice adapts DefiScan v1's Autonomy dimension; dependencies are one of several criteria, not the whole frame.

GUARDRAILS (read before grading):
- Category alone (Liquid Staking, Bridge, RWA Lending, Restaking, …) does NOT force a grade. A category is a hint about where to look; the grade must come from the concrete A1–A9 findings below.
- Base-chain consensus (Ethereum PoS, the chain's validator set, the canonical Deposit Contract at 0x00000000219ab540356cBB839Cbe05303d7705Fa) is the SUBSTRATE, not a dependency, for any protocol deployed on that chain. Do not list "depends on Ethereum" as a finding.
- Oracles or other integrations used by DOWNSTREAM protocols that happen to read this protocol's token (e.g. Chainlink stETH/USD consumed by Aave) are NOT this protocol's dependencies. Count only what THIS protocol's contracts call or trust on-chain. If a feed or contract appears in the protocol's docs only as reference material for third-party integrators, EXCLUDE IT ENTIRELY — do not log it as a finding even with a "peripheral" or "referenced only" caveat.
- "Upgradeable admin can change things" belongs to the CONTROL slice; only count it here when the upgrade surface lets governance silently swap an external dependency (see A9).
- Underlying-asset risk in opt-in, isolated markets is NOT autonomy-red on its own. When a protocol wraps third-party yield-bearing assets (LSTs, LRTs, ERC-4626 vaults, lending receipts, restaked tokens) into per-market silos that users explicitly choose, a failure of one underlying does NOT propagate to other markets and is risk the user opted into per-market. Record it under A4 with depth and propagation scope, but do not let it alone drive a red verdict — red requires an external dependency that cross-cuts the protocol or that the user did not opt into at deposit time. A failure mode that is "if the LRT you deposited is hacked, your principal in that LRT-backed market is impaired" is the underlying's autonomy story, not this protocol's; grade it on whether THIS protocol introduces additional dependencies on top.
- Sub-module enumeration is mandatory before grading. If the protocol ships distinct product lines or modules (e.g., a v2 core AMM plus a newer perps/funding-rate module, a lending pool plus a separate vault layer, an L1 core plus a cross-chain extension), enumerate each in findings and grade against the WORST module weighted by its share of TVS or its blast radius. A green core does not rescue an orange/red sub-module; conversely, a small red sub-module with capped TVS may bound the overall grade to orange. Name each module by its on-chain factory or router address. If you do not know whether a module exists, that is an unknowns[] entry, not silence. EXPLICIT TVS WEIGHTING: in the verdict, state each module's approximate share of total protocol TVS (use "~X%" if exact figures unavailable; check DeFiLlama or block-explorer balances on the module's main contract) and how that share informs the weighted grade. Format: "Module A holds ~X% of TVS (grade: <g>); Module B holds ~Y% (grade: <g>); weighted overall = <grade> because <reason>." If a red sub-module holds <5% of total TVS and is capped, the overall grade may be orange; if it holds >25%, the overall grade is red. Do not let qualitative reasoning substitute for the percentages — write the numbers.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. External contract calls. Enumerate every external contract the core contracts call or read from (oracles, price feeds, AMM pools, lending pools, staking/deposit contracts, yield wrappers). For each, identify the address, the provider, and what user-facing function of this protocol would break or mis-price if that external contract paused, mis-reported, or behaved adversarially. Grep for "oracle", "aggregator", "getPrice", "latestAnswer", "chainlink", "pyth", "redstone" as a starting point. To verify an oracle is live and what it currently reports, hit https://defipunkd.com/api/contract/read?chainId=<id>&address=<oracle>&method=latestAnswer (or &method=latestRoundData, &method=getPrice, &method=description; use the BARE method name without parens) and cite the URL — the response includes blockNumber/blockHash and rawReturnData, which is stronger evidence than a docs page about which feed "should" be used.
- A2. Off-chain actor committees reporting INTO the protocol. Oracle committees, guardian multisigs, DAO-selected validator sets acting as protocol reporters, exit-bus signers, fraud-proof challengers. For each, record committee size, quorum, who picks the members, and what mis-reporting could do (mint, burn, finalize withdrawal, freeze). Distinguish this from governance admins (control slice). NOTE ON STAKING PROTOCOLS: validator slashing and node-operator misbehavior are properties of the base-chain substrate, NOT external dependencies, when the operator set is diversified enough that a coordinated failure caps at <5% principal loss. Count validator/operator risk under A2 ONLY if the operator set is small, non-diversified, or lacks bonding / slashing-insurance / diversification mitigations. A curated set of 30+ independent operators with documented diversification falls under the mitigated path; a 3-operator LST does not. Do not cite the protocol's own risk disclosure as evidence that operator failure = principal loss unless you also check the diversification and bond mitigations.
- A3. Bridge / cross-chain messaging dependencies. Only count bridges that carry material TVL or are required for a core user flow. For each, name the bridge operator (canonical L1↔L2, LayerZero, Wormhole, Axelar, custom multisig), the trust model (canonical, optimistic, light-client, guardian set), and what fraction of TVL or users ride on it. "wstETH exists on 15 chains" is not a finding unless material TVL sits there. Before listing any non-primary chain deployment as a dependency, verify it is still operational as of analysis_date — retired or sunset deployments (e.g., Lido-on-Terra, Lido-on-Solana) belong in unknowns[] or should be omitted, never cited as a current dependency.
- A4. Nested collateral / restaking chains. For restaking / LRT / receipt-of-receipt designs: record the depth of the collateral chain, every actor with slashing or freezing power at each level, and whether a failure N levels deep propagates to user principal here.
- A5. Fork lineage (silent check). If DeFiLlama's forkedFrom is non-empty, record it as one finding and move on. If empty, do NOT add a placeholder finding; it adds noise.
- A6. Fallback mechanisms and circuit breakers. What catches an external failure? Sanity-check contracts on oracle reports, rebase bounds, pause paths triggered by bad prices, second-opinion oracles, max-per-block throughput caps, withdrawal queues that absorb bad reports. Record which A1–A4 risks are mitigated by which fallback, and which are unmitigated. For EACH fallback you cite, state its activation status explicitly: (i) LIVE and enforcing on-chain today, (ii) DEPLOYED but not yet wired / activated (e.g., interface exists but the address is zero or the role is unassigned), or (iii) DOCUMENTED / PROPOSED only (forum post, LIP draft, audit pending). Only (i) counts as mitigation for the grade. A fallback in state (ii) or (iii) should be noted but must not reduce the risk in your steel-man or verdict. If you cannot determine activation status, add an unknowns[] entry rather than assume it is live.
- A7. Sequencer / L1-liveness dependency BEYOND the base-chain substrate. SCOPE — sequencer risk only counts here when the protocol IS its own L2/L3 appchain or app-rollup, where the sequencer is part of the protocol's own stack and a freeze is a protocol-level outage. A protocol permissionlessly deployed on a third-party L2 inherits that L2's sequencer as substrate, not an A7 dependency. Record the sequencer/DA trust model when A7 applies.
- A8. Keeper / relayer / off-chain bot liveness. Protocols that need permissionless-but-necessary off-chain actors (liquidation bots, auto-compounders, deposit relayers, intent solvers). Record whether the role is permissionless, what degrades if nobody runs it (yield paused, bad debt accumulates, positions go stale), and whether the failure mode is graceful or catastrophic.
- A9. Governance-mutable dependency surface. Can an admin or DAO action silently INTRODUCE a NEW EXTERNAL dependency — swap the oracle address to a different provider, register a new staking module that calls an untrusted contract, add a new bridge, route SY through a new external vault — without an exit window for users? Check the upgrade / router / module-registry contracts. Answer: which EXTERNAL dependencies are governance-mutable, who holds that power, and whether there is a timelock or exit window. SCOPE LIMIT — read carefully: A9 is about the *external dependency surface*, not the upgrade surface in general. "The proxyAdmin / EOA can upgrade the router implementation to arbitrary bytecode" is a CONTROL-slice finding (admin can rug), NOT an autonomy-A9 finding. A9 fires only when the upgrade specifically swaps out or adds a contract that THIS protocol calls or trusts (e.g., changing the oracle address from Chainlink to a malicious feed, registering a new SY adapter that points to a third-party vault, redirecting a bridge endpoint). If the only finding is "admin can change implementation," do not log it under A9 and do not let it drive the autonomy grade — note it under control instead. The autonomy-relevant version of the same upgrade key is "admin can swap [specific external address X] without timelock"; that requires identifying the specific external dependency that becomes mutable.

STEEL-MAN (per Hard Rule 13): write one-sentence strongest arguments for red, orange, and green using the A1–A9 findings.

IMPACTED TVS ESTIMATE: the headline MUST include a rough impacted-TVS figure — the fraction of protocol TVS that could be lost or frozen if the worst-unmitigated dependency you identified failed. Use "~X%" if exact numbers are unavailable, "<1%" for de minimis, "unclear" only if A1–A9 left the question genuinely open (in which case grade=unknown is usually correct). Do NOT substitute qualitative phrases like "significant" — give a number or bracket.

GRADE ANCHORS (mapped to DefiScan v1 stages):
- green = Stage 2 equivalent. Failure of any external dependency cannot cause loss of user principal or unclaimed yield. Either there are no material external dependencies, or every critical one has a documented fallback (A6) that keeps users whole. Governance cannot silently introduce new dependencies without an exit window (A9). Impacted TVS under any single-dependency failure: effectively 0.
- orange = Stage 1 equivalent. Failure of some external dependency can cause loss of unclaimed yield, or can materially change expected performance (pause withdrawals, freeze positions, degrade price quality), but cannot cause loss of principal. Committee-based oracles with sanity checks, canonical-only bridges, fallback paths that exist but are incomplete, or governance-mutable dependencies protected by a ≥7-day timelock. Impacted TVS is bounded and recoverable.
- red = Stage 0 equivalent. Failure of an external dependency CAN cause theft or loss of principal. Examples: single-provider oracle with no sanity check or fallback, material TVL on a non-canonical bridge with a guardian multisig, governance can hot-swap oracles or add staking modules with no timelock or exit window, unmitigated keeper-liveness dependency where positions become insolvent if bots stop. Impacted TVS is material.
- unknown = checklist incomplete after inspecting source + verified contracts. Prefer unknown over guessing when A1/A6 could not be reconstructed. RESERVE unknown for cases where the CORE ARCHITECTURE itself is unverifiable — not for cases where you merely cannot enumerate every per-market dependency in a multi-market protocol. If the core router/factory/oracle architecture is verifiable on-chain and you can determine whether the core requires external dependencies, grade the architecture even when an exhaustive per-market external-dependency census is infeasible. Acknowledge the per-market gap in unknowns[] but still issue a grade. Refusing to grade a multi-market protocol because you cannot list every SY/vault/market is over-use of unknown; grade the architecture and say so.

PROMPT-META CHECK (per Hard Rule 17): before finalizing, verify the verdict cites concrete contract addresses, docs, or code — not the rubric itself. If your verdict says "the protocol belongs to a category the rubric marks red", rewrite it with the A1–A9 finding that actually justifies the grade, or drop to grade=unknown.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "autonomy",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Open Access Unverified Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-cow-amm
- protocol.name:              Balancer CoW AMM
- protocol.chains:            xDai, Ethereum, Arbitrum, Base
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: OPEN-ACCESS

Evaluate who is allowed to use the protocol and whether any of that permission is granted off-chain.

Scope: this slice is about ADMISSION — who can enter, exit, or transact. Operator LIVENESS (what breaks if keepers/oracles go offline) is assessed in the dependencies slice and is out of scope for the grade here. You may note operator dependencies as context, but do not let "the protocol halts if operator X disappears" drive the access grade on its own; that belongs in dependencies. Source verification / contract verification on block explorers is assessed in the verifiability slice and is out of scope here — do NOT let "contract is unverified" drive the access grade.

Framing: the smart contracts are the access layer; frontends are UX. A permissionless contract is reachable by any client (SDK, third-party UI, aggregator, wallet integration). Frontend ToS, IP geo-blocking, and wallet screening are publisher policies on one specific client — they are reported as context but do NOT determine the grade. The grade hinges on (1) what the contract itself permits, and (2) whether the protocol is practically reachable without the official publisher's cooperation.

Meta-check before finalizing: if your verdict cites phrases from this prompt as evidence ("the protocol meets the 'credible alternatives' condition", "this fits the 'documented fallback' rule"), redo the verdict. The prompt describes the rubric; evidence must come from the protocol. A verdict should cite what the protocol does, not what the rubric says.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. Whitelist / allowlist modifiers in user-facing entry points. Grep for "onlyWhitelisted", "onlyRole", "allowlist", "isAccredited", "isKYCed". Note which functions are gated and who can add/remove from the list.
- A2. Off-chain operators in the admission path: keepers, sequencers, privileged relayers, oracle posters whose approval is required to admit a user action (not just to keep the protocol live). For each, identify whether the role is held by a single operator, a permissioned committee, or is permissionless. Enumerate per user-facing function class (deposit vs withdraw-request vs claim-finalized vs transfer) which ones require operator approval to be admitted, and which ones admit users unconditionally. A function whose placement is unconditional but whose downstream settlement depends on an operator is an admission-permissionless function — flag the liveness dependency as context and defer its grading weight to the dependencies slice.
- A3. Frontend restrictions on the official interface — record as context, not as a grade lever. Distinguish:
    - A3-passive: boilerplate ToS clauses (sanctions attestation, restricted-territory self-certification, VPN-circumvention prohibition, "comply with applicable law" eligibility, age of majority).
    - A3-active: runtime enforcement — IP-based geo-blocking, wallet-address screening against a sanctions oracle (Chainalysis, TRM, Elliptic), KYC wall, rendering-blocking jurisdiction banner.
  Record findings under the correct tier. Quote ToS text or banner text in evidence[].shows. These findings populate the headline and rationale but do NOT move the grade by themselves; the grade is set by A1, A2, and the A3b path check below.
- A3b. Independent access paths (the operative grade input). Enumerate paths that do not require the official publisher's cooperation:
    - Published SDK / library / CLI for direct contract interaction.
    - Third-party frontends operated by separate legal entities.
    - Wallet-integrated access (MetaMask Swaps, Safe apps, etc.).
    - DEX / lending / yield aggregators that route through the contracts.
  Record at least one concrete link per path that exists. The protocol does NOT have to self-document these — the test is existence, not UX cost. An A3b-i redistribution of the official UI bound by the same ToS does NOT count as an independent path.
- A4. Sanctions / compliance tooling at the contract level: does the protocol check addresses against OFAC lists or similar on-chain blocklists in the contract itself? (Frontend-only screening belongs in A3.)
- A5. Differentiate read access vs write access: many protocols are read-permissionless (anyone can view state) but write-gated (only certain addresses can deposit/borrow). Record both.
- A6. ToS / Legal links: locate them on the website and produce a VERBATIM quote of any jurisdictional, sanctions, or eligibility clause in evidence[].shows. If you cannot extract the clause text verbatim (SPA render failure, paywall, dead link, etc.), do NOT paraphrase or infer from general knowledge — record the ToS URL in unknowns[] with the reason extraction failed. Assertions about ToS content without a verbatim quote will be downweighted by reviewers.

Then write the steel-man section per Hard Rule 11.

Grade rules (admission-focused; liveness concerns belong in dependencies; source verification belongs in verifiability):
- green   = no contract-level whitelist/KYC on user entry/exit; no operator approval required to admit a user action; AND at least one independent A3b path exists (published SDK, third-party frontend, wallet integration, or aggregator routing). Frontend ToS posture and A3-active enforcement on the official UI do NOT block green when contracts are permissionless and an independent path exists — they are reported as context.
- orange  = contracts admit users unconditionally, BUT the protocol is operationally captured by the official publisher: no published SDK, no third-party frontend, no wallet integration, no aggregator routing. The contract is theoretically open but practically reachable only through the official UI. Also applies when admission requires approval from a permissioned committee that is governance-managed with a documented replacement procedure.
- red     = contract-level whitelist / KYC on user entry/exit, OR admission of a core user action requires approval from a single privileged operator or a small committee with no documented replacement procedure, OR enforces an on-chain blocklist updatable by a single party.
- unknown = checklist incomplete after checking the sources above.

Default-grade guidance: when contracts are fully permissionless AND any A3b independent path exists, the default grade is green regardless of frontend ToS or A3-active enforcement on the official UI. Frontend geo-blocking, sanctions-oracle wallet screening, and ToS sanctions clauses are publisher policies on one client and are reported in findings/headline as context, not as grade levers. To grade orange on operational-capture grounds, the auditor must affirmatively show that ALL independent paths are absent or also gated.

Guideline on committees: where admission depends on a multi-operator committee, the relevant axes are (a) set size, (b) whether replacement/rotation is governed on-chain, (c) whether the replacement procedure is publicly documented. A large set with on-chain governance replacement should not be graded as a single-party operator even if rotation is not instantaneous. A small set with informal replacement should be treated as a single-party operator.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "open-access",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Audit all 5 dimensions · one prompt Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-cow-amm
- protocol.name:              Balancer CoW AMM
- protocol.chains:            xDai, Ethereum, Arbitrum, Base
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: CONTROL

Evaluate who can change the protocol's rules, how fast, and how broadly.

(Step 0 capability probe and the off-chain-only fallback live in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[])

- **C1.** For each address you assess: who is the contract owner / admin / pendingAdmin / governor — read these via the block explorer's "Read Contract" tab OR `https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=owner` (BARE method names: `&method=owner`, `&method=admin`, `&method=pendingOwner`, `&method=governor`). For Safes use `/api/safe/owners?chainId=<id>&address=0x...`. When a protocol has multiple major versions deployed (v2/v3/v4), perform C1 reads on the NEWEST deployment separately — newer deployments often have weaker control surfaces than the legacy core.
- **C2.** Upgrade mechanism: transparent proxy / UUPS / Beacon / Diamond / immutable. Identify the proxy admin address. Check upgradeability of GOVERNANCE contracts too — a Governor / Aragon Voting / OZ Governor is often itself a proxy whose admin is the Timelock. Asymmetry: when fund-holding cores are immutable AND governance has no admin path that reaches them, an upgradable Governor/Timelock is T3-only and must NOT drag the verdict below green on that basis alone (see grade rules and the "immutable cores" caveat). Only call upgradability "mixed" if you can name a concrete function on the upgradable surface that reaches T1 or T2 on user funds.
- **C3.** EXECUTION PATH (enumerate every stage, in order, with delays in seconds). The operative path is usually a chain (voting → scheduler → timelock → executor; or governor → queue → execute; or Aragon Voting → DualGovernance → EmergencyProtectedTimelock → AdminExecutor). For each stage, record (a) the contract address, (b) the delay constant name + value in seconds, (c) the URL you read it from (block-explorer Read Contract OR `/api/contract/read?...&method=MIN_DELAY&block=<n>`). Do NOT stop at the first timelock-shaped contract — if its admin is itself called by another contract, keep walking. The grading delay is the SUM OF DELAYS ON THE UNCONTESTED FAST PATH (shortest time a proposal with no opposition can go from submission to executable). Dynamic / contested extensions (veto signaling, rage quit, escrow delay) are modifiers, not the basis — note them separately.
- **C4.** Enumerate EVERY multisig with reachable control — main proxy admin, emergency activation, emergency execution, reseal / pause, gate-seal committees, tiebreaker, per-module admins. For each Safe, fetch threshold + owners + version via `/api/safe/owners?chainId=<id>&address=0x...` (response includes raw eth_call data, so the URL is citable evidence). Enumerate ops/council/incentives multisigs even when off the upgrade path — record their scope so a reader can see they are NOT on the upgrade path. For each: (a) address, (b) threshold / total signers, (c) signer identities classified as insider (team, paid auditors under ongoing engagement, mandated service providers) vs non-insider (independent community members, unaffiliated researchers), (d) the specific power held (upgrade, pause, parameter, etc.).
- **C5.** On-chain governance: Governor / GovernorBravo / OZ Governor / Aragon Voting with token-weighted voting? Record proposal threshold, voting period, quorum, and the timelock delay between queue and execute. Every numeric constant must come from a Read Contract call you can link to, or be in unknowns[] with the C-code. If votingDelay / votingPeriod are denominated in BLOCKS, convert to seconds at the chain's CURRENT block time (Ethereum mainnet ≈ 12s post-Merge, not the 15s in older Compound/Bravo deployments) — cite both block count and converted seconds.
- **C6.** EMERGENCY POWERS: separate emergency-pause / guardian role with a different time cap or different actor than the main upgrade authority? Record it explicitly.
- **C7.** POWER TIER (blast radius). For each privileged path in C3–C6, classify the WORST thing that path can do, choosing the highest applicable tier. Cite the specific function name and any on-chain bound — tier claims without a named function are unsupported.
    - **T1 — FUND-CRITICAL**: replace implementation of contracts holding user funds; change AMM math / accounting / collateral logic; mint unbacked debt or shares; pause withdrawals; drain user-fund treasury; change oracle to attacker-controlled source; replace upgrade admin with EOA.
    - **T2 — ECONOMICALLY MATERIAL**: change fee parameters within bounded ranges; redirect protocol fees; add/remove markets / collateral types; bounded inflation or token mint within hard-capped schedule; spend protocol-owned (non-user) treasury.
    - **T3 — GOVERNANCE-INTERNAL**: change voting rules, quorum, voting period, proposal threshold; upgrade the Governor itself; rotate Timelock admin; mint governance tokens within a capped annual schedule.
    - **T4 — OPERATIONAL**: incentives distribution, grants, ENS / frontend canonicalization, deployment coordination, periphery router deprecation.
  The grade is set by the HIGHEST tier reachable on the uncontested fast path, not the median. State the tier and the binding function in the verdict.

### Read Contract discipline (applies to C3, C4, C5)

Every numeric constant cited (timelock delays, voting periods, multisig thresholds, quorum percentages) must come from EITHER (a) a block-explorer Read Contract URL, OR (b) a DeFiPunkd `/api/contract/read` or `/api/safe/owners` URL (preferred with `&block=<n>` for content-addressed evidence), OR (c) an unknowns[] entry with the C-code. Docs / blog posts are corroboration only — they cannot be the sole citation for a value that is also readable on-chain.

### Off-chain-only substitute hierarchy (when grading_basis="off-chain-only" — see preamble Rule 16)

When on-chain reads were genuinely unreachable this run, eligible off-chain substitutes in priority order:
1. Linked audit PDFs (admin roles, multisig members, timelock delays usually enumerated).
2. Governance forum posts that quote constants from a successful on-chain proposal (cite post URL + linked execution-tx URL).
3. Official protocol docs pages with named addresses and roles (must be on a domain owned by the protocol).
4. GitHub README / SECURITY.md / governance/*.md at a pinned commit SHA.

Forbidden substitutes: third-party blog posts, X / Twitter threads, search-result snippets, model memory. Required degradation: any C-code citing a numeric constant from docs/forum/audit prose ONLY must also carry an `unknowns[]` entry with `-offchain` suffix noting "value not re-read on-chain in this run; corroboration only".

### Grade rules (apply the timelock bar conditional on the highest C7 tier reachable on the fast path)

Security Council standard (used below): a multisig qualifies as "Security Council" only if ALL of: ≥7 signers, ≥51% threshold, ≥50% non-insider signers, every signer publicly announced. Failing any criterion = NOT a Security Council, regardless of signer reputation.

- **green**: highest reachable tier is T3 or T4 regardless of timelock; OR T2 reachable with uncontested-fast-path delay ≥7 days; OR T1 reachable only via immutable contracts (T1 is unreachable); OR T1 reachable with uncontested-fast-path delay ≥7 days combined with a Security Council multisig; OR T1 reachable with uncontested-fast-path delay ≥7 days through active on-chain governance with broad token distribution.
- **orange**: T2 reachable with uncontested-fast-path delay >0 but <7 days; OR T1 reachable with uncontested-fast-path delay >0 but <7 days; OR a multisig failing one or more Security Council criteria sits on a T1/T2 path; OR unclear upgrade authority on a T1/T2 path; OR governance with very short timelock or low quorum on a T1/T2 path.
- **red**: T1 reachable with no timelock by a single EOA or 2-of-3 multisig; OR a T1 upgrade admin that is not a smart contract you can audit.
- **unknown**: completed the checklist but still cannot determine the upgrade authority OR cannot classify the highest tier reachable on the main contracts.

Tiering caveats:
- "Bounded" must be enforced ON-CHAIN to count as T2. A function that sets fees with no upper-bound check is T1 — cite the bound check.
- Recurring T2 economic extraction (e.g. fee redirect with no rate limit) approaches T1 over time. A single proposal that can permanently redirect all future revenue is T1.
- T3 assumes the governance contract cannot itself authorize a T1/T2 action without going through the same timelock. If governance can self-upgrade to bypass the timelock, T3 collapses into T1.
- Do not downgrade tier by hand-waving ("realistically governance would never…"). Tier on what the contract permits, not what feels likely.

Notes:
- **Dynamic / dual-governance timelocks** (Lido, Compound escrow veto): the rubric grades on the uncontested path because that is the path most upgrades take. A dynamic extension that fires only under stake-weighted opposition is a real protection — name it in the green steel-man, but it does not lift an orange fast path into green; state the tension in the verdict.
- **Immutable cores with upgradable governance** (Uniswap-style): if fund-holding contracts are immutable and have no admin-reachable function moving / freezing / re-routing user funds, the highest reachable tier on the upgrade path is T3 — green regardless of timelock. Don't grade this orange just because the Governor is a proxy — that's a C2 fact, not a downgrade. Downgrade only applies if you can cite a concrete function on the upgradable surface that reaches T1 or T2 (privileged hook, upgradable factory controlling fund-routing, fee-switch redirecting protocol revenue without bound).
- **The 7-day bar** reflects the exit-window standard — users need notice after a queued upgrade to withdraw if they disagree. The ability-to-exit slice grades the exit side; this slice grades the delay side; both must hold for users to actually benefit from the delay.

---

### Slice: ABILITY-TO-EXIT

Evaluate whether users can withdraw their funds on their own terms, even under adversarial admin conditions.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- E1. Enumerate every user-facing exit function in the main contracts: withdraw, redeem, burn, requestWithdrawal, claim, exit, etc. List them by name. Do NOT treat the contract as a monolith.
- E2. For EACH exit function in E1: identify its access modifiers and any pause guards (e.g. _checkResumed, whenNotPaused, onlyRole). Functions that gate REQUEST PLACEMENT often differ from functions that CLAIM FINALIZED FUNDS — check both separately.
- E3. For each pause guard: identify the role holder (which address holds PAUSE_ROLE / GUARDIAN / etc.) and the maximum pause duration. Specifically check whether PAUSE_INFINITELY (or equivalent uncapped pause) is callable, and which actor can call it (single multisig vs governance vote). For role-holder reads use https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=hasRole&args=0x...,0x... or &method=getRoleAdmin&args=0x.... For "is currently paused" checks use &method=paused or &method=isPaused&args=<resume-code>. Use the BARE method name (no parens). Cite the URL with &block=<n> in evidence[].
- E4. EMERGENCY vs GOVERNANCE pause distinction: many protocols have a fast-acting emergency pause capped at N days and a slow governance pause that can be indefinite. Record both paths separately if present, with their time caps and actor classes.
- E5. Queued redemption: documented maximum queue duration, daily withdrawal caps, whether the queue itself is pausable.
- E6. Forced-exit / escape-hatch / permissionless emergency-exit mechanism for adversarial-admin scenarios.
- E7. Frontend dependency: confirm exit functions are directly callable on-chain (e.g. via Etherscan write tab or a generic wallet) without the project's frontend.

Then write the steel-man section per Hard Rule 11. Common red-vs-orange tension on this slice: indefinite pause exists (suggests red) BUT the realistic emergency path is time-capped AND claims of already-finalized exits are not pause-gated (suggests orange). Resolve this by stating who can do what for how long, not by stopping at the worst-case sentence.

Grade rules:
- green   = permissionless exit; pause is either absent, narrowly scoped to clearly-described emergencies with auto-expiry, or capped at ≤7 days; no frontend dependency for exit; claims of already-finalized exits are not pause-gated under any path.
- orange  = pausable with broad scope OR indefinite pause is reachable only through governance vote (not unilateral admin action), OR queued redemption with documented max > 7 days, OR claims-of-finalized are exempt but new-request placement can be paused indefinitely by governance.
- red     = exit requires admin signature, OR ANY actor (including governance) can pause CLAIMS of finalized exits indefinitely, OR there is no on-chain exit function at all (purely custodial), OR pause is held by a single EOA / 2-of-3 multisig with no time cap.
- unknown = checklist incomplete after checking the sources above.

---

### Slice: AUTONOMY

Evaluate this protocol's autonomy: can a failure of anything outside its own contracts cause theft or loss of user principal, loss of unclaimed yield, or materially change the protocol's expected performance? "Autonomy" is not "has zero external touchpoints" — it is "external failures are either survivable or recoverable without user loss". This slice adapts DefiScan v1's Autonomy dimension; dependencies are one of several criteria, not the whole frame.

GUARDRAILS (read before grading):
- Category alone (Liquid Staking, Bridge, RWA Lending, Restaking, …) does NOT force a grade. A category is a hint about where to look; the grade must come from the concrete A1–A9 findings below.
- Base-chain consensus (Ethereum PoS, the chain's validator set, the canonical Deposit Contract at 0x00000000219ab540356cBB839Cbe05303d7705Fa) is the SUBSTRATE, not a dependency, for any protocol deployed on that chain. Do not list "depends on Ethereum" as a finding.
- Oracles or other integrations used by DOWNSTREAM protocols that happen to read this protocol's token (e.g. Chainlink stETH/USD consumed by Aave) are NOT this protocol's dependencies. Count only what THIS protocol's contracts call or trust on-chain. If a feed or contract appears in the protocol's docs only as reference material for third-party integrators, EXCLUDE IT ENTIRELY — do not log it as a finding even with a "peripheral" or "referenced only" caveat.
- "Upgradeable admin can change things" belongs to the CONTROL slice; only count it here when the upgrade surface lets governance silently swap an external dependency (see A9).
- Underlying-asset risk in opt-in, isolated markets is NOT autonomy-red on its own. When a protocol wraps third-party yield-bearing assets (LSTs, LRTs, ERC-4626 vaults, lending receipts, restaked tokens) into per-market silos that users explicitly choose, a failure of one underlying does NOT propagate to other markets and is risk the user opted into per-market. Record it under A4 with depth and propagation scope, but do not let it alone drive a red verdict — red requires an external dependency that cross-cuts the protocol or that the user did not opt into at deposit time. A failure mode that is "if the LRT you deposited is hacked, your principal in that LRT-backed market is impaired" is the underlying's autonomy story, not this protocol's; grade it on whether THIS protocol introduces additional dependencies on top.
- Sub-module enumeration is mandatory before grading. If the protocol ships distinct product lines or modules (e.g., a v2 core AMM plus a newer perps/funding-rate module, a lending pool plus a separate vault layer, an L1 core plus a cross-chain extension), enumerate each in findings and grade against the WORST module weighted by its share of TVS or its blast radius. A green core does not rescue an orange/red sub-module; conversely, a small red sub-module with capped TVS may bound the overall grade to orange. Name each module by its on-chain factory or router address. If you do not know whether a module exists, that is an unknowns[] entry, not silence. EXPLICIT TVS WEIGHTING: in the verdict, state each module's approximate share of total protocol TVS (use "~X%" if exact figures unavailable; check DeFiLlama or block-explorer balances on the module's main contract) and how that share informs the weighted grade. Format: "Module A holds ~X% of TVS (grade: <g>); Module B holds ~Y% (grade: <g>); weighted overall = <grade> because <reason>." If a red sub-module holds <5% of total TVS and is capped, the overall grade may be orange; if it holds >25%, the overall grade is red. Do not let qualitative reasoning substitute for the percentages — write the numbers.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. External contract calls. Enumerate every external contract the core contracts call or read from (oracles, price feeds, AMM pools, lending pools, staking/deposit contracts, yield wrappers). For each, identify the address, the provider, and what user-facing function of this protocol would break or mis-price if that external contract paused, mis-reported, or behaved adversarially. Grep for "oracle", "aggregator", "getPrice", "latestAnswer", "chainlink", "pyth", "redstone" as a starting point. To verify an oracle is live and what it currently reports, hit https://defipunkd.com/api/contract/read?chainId=<id>&address=<oracle>&method=latestAnswer (or &method=latestRoundData, &method=getPrice, &method=description; use the BARE method name without parens) and cite the URL — the response includes blockNumber/blockHash and rawReturnData, which is stronger evidence than a docs page about which feed "should" be used.
- A2. Off-chain actor committees reporting INTO the protocol. Oracle committees, guardian multisigs, DAO-selected validator sets acting as protocol reporters, exit-bus signers, fraud-proof challengers. For each, record committee size, quorum, who picks the members, and what mis-reporting could do (mint, burn, finalize withdrawal, freeze). Distinguish this from governance admins (control slice). NOTE ON STAKING PROTOCOLS: validator slashing and node-operator misbehavior are properties of the base-chain substrate, NOT external dependencies, when the operator set is diversified enough that a coordinated failure caps at <5% principal loss. Count validator/operator risk under A2 ONLY if the operator set is small, non-diversified, or lacks bonding / slashing-insurance / diversification mitigations. A curated set of 30+ independent operators with documented diversification falls under the mitigated path; a 3-operator LST does not. Do not cite the protocol's own risk disclosure as evidence that operator failure = principal loss unless you also check the diversification and bond mitigations.
- A3. Bridge / cross-chain messaging dependencies. Only count bridges that carry material TVL or are required for a core user flow. For each, name the bridge operator (canonical L1↔L2, LayerZero, Wormhole, Axelar, custom multisig), the trust model (canonical, optimistic, light-client, guardian set), and what fraction of TVL or users ride on it. "wstETH exists on 15 chains" is not a finding unless material TVL sits there. Before listing any non-primary chain deployment as a dependency, verify it is still operational as of analysis_date — retired or sunset deployments (e.g., Lido-on-Terra, Lido-on-Solana) belong in unknowns[] or should be omitted, never cited as a current dependency.
- A4. Nested collateral / restaking chains. For restaking / LRT / receipt-of-receipt designs: record the depth of the collateral chain, every actor with slashing or freezing power at each level, and whether a failure N levels deep propagates to user principal here.
- A5. Fork lineage (silent check). If DeFiLlama's forkedFrom is non-empty, record it as one finding and move on. If empty, do NOT add a placeholder finding; it adds noise.
- A6. Fallback mechanisms and circuit breakers. What catches an external failure? Sanity-check contracts on oracle reports, rebase bounds, pause paths triggered by bad prices, second-opinion oracles, max-per-block throughput caps, withdrawal queues that absorb bad reports. Record which A1–A4 risks are mitigated by which fallback, and which are unmitigated. For EACH fallback you cite, state its activation status explicitly: (i) LIVE and enforcing on-chain today, (ii) DEPLOYED but not yet wired / activated (e.g., interface exists but the address is zero or the role is unassigned), or (iii) DOCUMENTED / PROPOSED only (forum post, LIP draft, audit pending). Only (i) counts as mitigation for the grade. A fallback in state (ii) or (iii) should be noted but must not reduce the risk in your steel-man or verdict. If you cannot determine activation status, add an unknowns[] entry rather than assume it is live.
- A7. Sequencer / L1-liveness dependency BEYOND the base-chain substrate. SCOPE — sequencer risk only counts here when the protocol IS its own L2/L3 appchain or app-rollup, where the sequencer is part of the protocol's own stack and a freeze is a protocol-level outage. A protocol permissionlessly deployed on a third-party L2 inherits that L2's sequencer as substrate, not an A7 dependency. Record the sequencer/DA trust model when A7 applies.
- A8. Keeper / relayer / off-chain bot liveness. Protocols that need permissionless-but-necessary off-chain actors (liquidation bots, auto-compounders, deposit relayers, intent solvers). Record whether the role is permissionless, what degrades if nobody runs it (yield paused, bad debt accumulates, positions go stale), and whether the failure mode is graceful or catastrophic.
- A9. Governance-mutable dependency surface. Can an admin or DAO action silently INTRODUCE a NEW EXTERNAL dependency — swap the oracle address to a different provider, register a new staking module that calls an untrusted contract, add a new bridge, route SY through a new external vault — without an exit window for users? Check the upgrade / router / module-registry contracts. Answer: which EXTERNAL dependencies are governance-mutable, who holds that power, and whether there is a timelock or exit window. SCOPE LIMIT — read carefully: A9 is about the *external dependency surface*, not the upgrade surface in general. "The proxyAdmin / EOA can upgrade the router implementation to arbitrary bytecode" is a CONTROL-slice finding (admin can rug), NOT an autonomy-A9 finding. A9 fires only when the upgrade specifically swaps out or adds a contract that THIS protocol calls or trusts (e.g., changing the oracle address from Chainlink to a malicious feed, registering a new SY adapter that points to a third-party vault, redirecting a bridge endpoint). If the only finding is "admin can change implementation," do not log it under A9 and do not let it drive the autonomy grade — note it under control instead. The autonomy-relevant version of the same upgrade key is "admin can swap [specific external address X] without timelock"; that requires identifying the specific external dependency that becomes mutable.

STEEL-MAN (per Hard Rule 13): write one-sentence strongest arguments for red, orange, and green using the A1–A9 findings.

IMPACTED TVS ESTIMATE: the headline MUST include a rough impacted-TVS figure — the fraction of protocol TVS that could be lost or frozen if the worst-unmitigated dependency you identified failed. Use "~X%" if exact numbers are unavailable, "<1%" for de minimis, "unclear" only if A1–A9 left the question genuinely open (in which case grade=unknown is usually correct). Do NOT substitute qualitative phrases like "significant" — give a number or bracket.

GRADE ANCHORS (mapped to DefiScan v1 stages):
- green = Stage 2 equivalent. Failure of any external dependency cannot cause loss of user principal or unclaimed yield. Either there are no material external dependencies, or every critical one has a documented fallback (A6) that keeps users whole. Governance cannot silently introduce new dependencies without an exit window (A9). Impacted TVS under any single-dependency failure: effectively 0.
- orange = Stage 1 equivalent. Failure of some external dependency can cause loss of unclaimed yield, or can materially change expected performance (pause withdrawals, freeze positions, degrade price quality), but cannot cause loss of principal. Committee-based oracles with sanity checks, canonical-only bridges, fallback paths that exist but are incomplete, or governance-mutable dependencies protected by a ≥7-day timelock. Impacted TVS is bounded and recoverable.
- red = Stage 0 equivalent. Failure of an external dependency CAN cause theft or loss of principal. Examples: single-provider oracle with no sanity check or fallback, material TVL on a non-canonical bridge with a guardian multisig, governance can hot-swap oracles or add staking modules with no timelock or exit window, unmitigated keeper-liveness dependency where positions become insolvent if bots stop. Impacted TVS is material.
- unknown = checklist incomplete after inspecting source + verified contracts. Prefer unknown over guessing when A1/A6 could not be reconstructed. RESERVE unknown for cases where the CORE ARCHITECTURE itself is unverifiable — not for cases where you merely cannot enumerate every per-market dependency in a multi-market protocol. If the core router/factory/oracle architecture is verifiable on-chain and you can determine whether the core requires external dependencies, grade the architecture even when an exhaustive per-market external-dependency census is infeasible. Acknowledge the per-market gap in unknowns[] but still issue a grade. Refusing to grade a multi-market protocol because you cannot list every SY/vault/market is over-use of unknown; grade the architecture and say so.

PROMPT-META CHECK (per Hard Rule 17): before finalizing, verify the verdict cites concrete contract addresses, docs, or code — not the rubric itself. If your verdict says "the protocol belongs to a category the rubric marks red", rewrite it with the A1–A9 finding that actually justifies the grade, or drop to grade=unknown.

---

### Slice: OPEN-ACCESS

Evaluate who is allowed to use the protocol and whether any of that permission is granted off-chain.

Scope: this slice is about ADMISSION — who can enter, exit, or transact. Operator LIVENESS (what breaks if keepers/oracles go offline) is assessed in the dependencies slice and is out of scope for the grade here. You may note operator dependencies as context, but do not let "the protocol halts if operator X disappears" drive the access grade on its own; that belongs in dependencies. Source verification / contract verification on block explorers is assessed in the verifiability slice and is out of scope here — do NOT let "contract is unverified" drive the access grade.

Framing: the smart contracts are the access layer; frontends are UX. A permissionless contract is reachable by any client (SDK, third-party UI, aggregator, wallet integration). Frontend ToS, IP geo-blocking, and wallet screening are publisher policies on one specific client — they are reported as context but do NOT determine the grade. The grade hinges on (1) what the contract itself permits, and (2) whether the protocol is practically reachable without the official publisher's cooperation.

Meta-check before finalizing: if your verdict cites phrases from this prompt as evidence ("the protocol meets the 'credible alternatives' condition", "this fits the 'documented fallback' rule"), redo the verdict. The prompt describes the rubric; evidence must come from the protocol. A verdict should cite what the protocol does, not what the rubric says.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. Whitelist / allowlist modifiers in user-facing entry points. Grep for "onlyWhitelisted", "onlyRole", "allowlist", "isAccredited", "isKYCed". Note which functions are gated and who can add/remove from the list.
- A2. Off-chain operators in the admission path: keepers, sequencers, privileged relayers, oracle posters whose approval is required to admit a user action (not just to keep the protocol live). For each, identify whether the role is held by a single operator, a permissioned committee, or is permissionless. Enumerate per user-facing function class (deposit vs withdraw-request vs claim-finalized vs transfer) which ones require operator approval to be admitted, and which ones admit users unconditionally. A function whose placement is unconditional but whose downstream settlement depends on an operator is an admission-permissionless function — flag the liveness dependency as context and defer its grading weight to the dependencies slice.
- A3. Frontend restrictions on the official interface — record as context, not as a grade lever. Distinguish:
    - A3-passive: boilerplate ToS clauses (sanctions attestation, restricted-territory self-certification, VPN-circumvention prohibition, "comply with applicable law" eligibility, age of majority).
    - A3-active: runtime enforcement — IP-based geo-blocking, wallet-address screening against a sanctions oracle (Chainalysis, TRM, Elliptic), KYC wall, rendering-blocking jurisdiction banner.
  Record findings under the correct tier. Quote ToS text or banner text in evidence[].shows. These findings populate the headline and rationale but do NOT move the grade by themselves; the grade is set by A1, A2, and the A3b path check below.
- A3b. Independent access paths (the operative grade input). Enumerate paths that do not require the official publisher's cooperation:
    - Published SDK / library / CLI for direct contract interaction.
    - Third-party frontends operated by separate legal entities.
    - Wallet-integrated access (MetaMask Swaps, Safe apps, etc.).
    - DEX / lending / yield aggregators that route through the contracts.
  Record at least one concrete link per path that exists. The protocol does NOT have to self-document these — the test is existence, not UX cost. An A3b-i redistribution of the official UI bound by the same ToS does NOT count as an independent path.
- A4. Sanctions / compliance tooling at the contract level: does the protocol check addresses against OFAC lists or similar on-chain blocklists in the contract itself? (Frontend-only screening belongs in A3.)
- A5. Differentiate read access vs write access: many protocols are read-permissionless (anyone can view state) but write-gated (only certain addresses can deposit/borrow). Record both.
- A6. ToS / Legal links: locate them on the website and produce a VERBATIM quote of any jurisdictional, sanctions, or eligibility clause in evidence[].shows. If you cannot extract the clause text verbatim (SPA render failure, paywall, dead link, etc.), do NOT paraphrase or infer from general knowledge — record the ToS URL in unknowns[] with the reason extraction failed. Assertions about ToS content without a verbatim quote will be downweighted by reviewers.

Then write the steel-man section per Hard Rule 11.

Grade rules (admission-focused; liveness concerns belong in dependencies; source verification belongs in verifiability):
- green   = no contract-level whitelist/KYC on user entry/exit; no operator approval required to admit a user action; AND at least one independent A3b path exists (published SDK, third-party frontend, wallet integration, or aggregator routing). Frontend ToS posture and A3-active enforcement on the official UI do NOT block green when contracts are permissionless and an independent path exists — they are reported as context.
- orange  = contracts admit users unconditionally, BUT the protocol is operationally captured by the official publisher: no published SDK, no third-party frontend, no wallet integration, no aggregator routing. The contract is theoretically open but practically reachable only through the official UI. Also applies when admission requires approval from a permissioned committee that is governance-managed with a documented replacement procedure.
- red     = contract-level whitelist / KYC on user entry/exit, OR admission of a core user action requires approval from a single privileged operator or a small committee with no documented replacement procedure, OR enforces an on-chain blocklist updatable by a single party.
- unknown = checklist incomplete after checking the sources above.

Default-grade guidance: when contracts are fully permissionless AND any A3b independent path exists, the default grade is green regardless of frontend ToS or A3-active enforcement on the official UI. Frontend geo-blocking, sanctions-oracle wallet screening, and ToS sanctions clauses are publisher policies on one client and are reported in findings/headline as context, not as grade levers. To grade orange on operational-capture grounds, the auditor must affirmatively show that ALL independent paths are absent or also gated.

Guideline on committees: where admission depends on a multi-operator committee, the relevant axes are (a) set size, (b) whether replacement/rotation is governed on-chain, (c) whether the replacement procedure is publicly documented. A large set with on-chain governance replacement should not be graded as a single-party operator even if rotation is not instantaneous. A small set with informal replacement should be treated as a single-party operator.

---

### Slice: VERIFIABILITY

Evaluate whether an outsider can independently confirm what the deployed code does.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- V1. For each address you assess: is the bytecode verified on the chain's block explorer? Record the "Contract Source Code Verified" indicator. https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x... returns this as a top-level "verified" boolean plus "abiSource" ("etherscan" / "sourcify") and an inline ABI — useful when the explorer page is rate-limited. If the contract is a proxy, verify BOTH the proxy contract AND the current implementation contract. The same /api/contract/abi response auto-resolves proxies and includes a "proxy.implementation" address when present, so one call covers V1 + V6 in one shot. An explorer "Similar Match" on a well-known proxy pattern (Aragon AppProxyUpgradeable, ERC1967Proxy, OssifiableProxy, OZ TransparentUpgradeableProxy) is expected for that pattern and does NOT count as a verification gap on its own — what matters is that the implementation is independently verified.
- V2. Source-to-repo correspondence: for each verified contract, attempt to find a matching commit in the linked GitHub repos. Record evidence[].commit on a match. Independent compile/bytecode-match is NOT required for green — a recognized public repo whose structure and file contents correspond to the explorer-visible source is sufficient. If you did not pin a commit SHA or run a bytecode diff, record that plainly in unknowns[] and proceed; it is a scope limit, not a downgrade signal.
- V3. Audit coverage: for each URL in protocol.audit_links, open it and record: auditor name, audit date, the specific contracts / commit in scope. Flag audits that predate the current deployment by >6 months without a follow-up review.
- V4. Auditor recognition: the following firms are broadly recognized in Solidity: Trail of Bits, Zellic, Spearbit, OpenZeppelin, ConsenSys Diligence, Certora (formal verification), Quantstamp, Halborn, Peckshield, Sigma Prime, ChainSecurity, Ackee Blockchain, MixBytes, Statemind. Unknown firms are orange-at-best for any green-grade claim. Name the firm explicitly in evidence[].
- V5. Post-audit drift: compare the most recent in-scope audit(s) against the currently-deployed source, weighted by what each contract does and by what the changes actually contain. SCOPE — drift only downgrades the grade when ALL of the following hold: (i) the drifting contract is fund-custody / settlement / accounting-critical (NOT a peripheral router, lens, quoter, or pure-view contract that holds no balances); (ii) the changes are material — new functions, modified access control, modified accounting, modified fund flow — and not refactors / struct relocations / import reorgs / build or CI fixes / formatting; (iii) no later audit, fix-audit, or differential audit from a recognized firm covers the changed files (audits often pin a pre-fix commit while a follow-up reviews the delta — match by file scope, not by commit-hash equality). When you cite drift as a downgrade reason, name the specific behavior change (function added, role granted, accounting formula altered) — "N commits ahead" or "+X/-Y LOC" alone is not evidence of material drift. If you have not sampled the diff content (e.g. via the GitHub compare view or the top commits in the window), record drift as an unknowns[] entry rather than auto-downgrading; commit-count and LOC are starting signals, not findings.
- V6. Implementation vs proxy: a verified proxy with an unverified implementation is effectively unverified. State whether the implementation is verified separately.

EVIDENCE DISCIPLINE (read before writing findings[]):
- Do not assert a specific deploy-commit SHA, bytecode equivalence, or "identical to audited commit" unless you actually fetched the artifact that shows it (e.g., a deployed-addresses JSON you opened, an explorer page you read). Inferred or plausible matches belong in unknowns[], never in findings[] or evidence[].
- Evidence[] entries must correspond to pages/files you actually retrieved this run. A URL you did not open is not evidence.

Then write the steel-man section per Hard Rule 11.

Grade rules:
- green   = deployed bytecode verified on the explorer (proxy AND implementation if proxied; "Similar Match" on a standard proxy pattern is fine per V1), a public source repo exists whose contents correspond to the explorer-visible source, AND ≥1 audit from a recognized firm covering the currently-deployed contracts (≤6 months of drift OR drift was re-audited). A missing local compile-match is not a downgrade — record it in unknowns[] and still grade green if the other conditions hold.
- orange  = verified but with visible drift from the public repo, OR audit scope is stale relative to deployment, OR only minor / unknown-firm audits exist, OR only some of the main contracts are verified, OR proxy verified but implementation only partially verified.
- red     = unverified bytecode (or verified proxy with unverified implementation), OR no audit in protocol.audit_links, OR no public repo.
- unknown = reserved for when the protocol's verifiability posture genuinely cannot be assessed (e.g., explorer and repo both inaccessible for this protocol). Do NOT use unknown merely because you, the analyst, could not run a particular check such as a bytecode diff — that goes in unknowns[] while the grade is still assigned from the evidence you do have.

---

### JSON output contract

Return exactly one JSON array inside a single ```json fenced block. The array MUST contain exactly five objects, one per risk slice, in this exact order: "control", "ability-to-exit", "autonomy", "open-access", "verifiability". Do not include the discovery slice.

Each object has the same shape as a normal slice submission:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "<one of: control | ability-to-exit | autonomy | open-access | verifiability>",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- Produce one complete object for each of these slices only: control, ability-to-exit, autonomy, open-access, verifiability.
- Reuse the same model, chat_url, snapshot_generated_at, prompt_version, analysis_date, and slug values across all five objects.
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches that object's slice checklist prefix verbatim (C1, E2, AU3, A3b, V4a, …); unknowns[] entries are checklist-coded ("C3: …").
- Wrap the array in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Verifiability tentative

Open source + 3 audits

Protocol publishes a GitHub repository and has at least one audit on record. This is a coarse Phase-0 signal only: auditor reputation, scope, and post-audit review coverage are not yet weighted.

Run your own prompt Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-cow-amm
- protocol.name:              Balancer CoW AMM
- protocol.chains:            xDai, Ethereum, Arbitrum, Base
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: VERIFIABILITY

Evaluate whether an outsider can independently confirm what the deployed code does.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- V1. For each address you assess: is the bytecode verified on the chain's block explorer? Record the "Contract Source Code Verified" indicator. https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x... returns this as a top-level "verified" boolean plus "abiSource" ("etherscan" / "sourcify") and an inline ABI — useful when the explorer page is rate-limited. If the contract is a proxy, verify BOTH the proxy contract AND the current implementation contract. The same /api/contract/abi response auto-resolves proxies and includes a "proxy.implementation" address when present, so one call covers V1 + V6 in one shot. An explorer "Similar Match" on a well-known proxy pattern (Aragon AppProxyUpgradeable, ERC1967Proxy, OssifiableProxy, OZ TransparentUpgradeableProxy) is expected for that pattern and does NOT count as a verification gap on its own — what matters is that the implementation is independently verified.
- V2. Source-to-repo correspondence: for each verified contract, attempt to find a matching commit in the linked GitHub repos. Record evidence[].commit on a match. Independent compile/bytecode-match is NOT required for green — a recognized public repo whose structure and file contents correspond to the explorer-visible source is sufficient. If you did not pin a commit SHA or run a bytecode diff, record that plainly in unknowns[] and proceed; it is a scope limit, not a downgrade signal.
- V3. Audit coverage: for each URL in protocol.audit_links, open it and record: auditor name, audit date, the specific contracts / commit in scope. Flag audits that predate the current deployment by >6 months without a follow-up review.
- V4. Auditor recognition: the following firms are broadly recognized in Solidity: Trail of Bits, Zellic, Spearbit, OpenZeppelin, ConsenSys Diligence, Certora (formal verification), Quantstamp, Halborn, Peckshield, Sigma Prime, ChainSecurity, Ackee Blockchain, MixBytes, Statemind. Unknown firms are orange-at-best for any green-grade claim. Name the firm explicitly in evidence[].
- V5. Post-audit drift: compare the most recent in-scope audit(s) against the currently-deployed source, weighted by what each contract does and by what the changes actually contain. SCOPE — drift only downgrades the grade when ALL of the following hold: (i) the drifting contract is fund-custody / settlement / accounting-critical (NOT a peripheral router, lens, quoter, or pure-view contract that holds no balances); (ii) the changes are material — new functions, modified access control, modified accounting, modified fund flow — and not refactors / struct relocations / import reorgs / build or CI fixes / formatting; (iii) no later audit, fix-audit, or differential audit from a recognized firm covers the changed files (audits often pin a pre-fix commit while a follow-up reviews the delta — match by file scope, not by commit-hash equality). When you cite drift as a downgrade reason, name the specific behavior change (function added, role granted, accounting formula altered) — "N commits ahead" or "+X/-Y LOC" alone is not evidence of material drift. If you have not sampled the diff content (e.g. via the GitHub compare view or the top commits in the window), record drift as an unknowns[] entry rather than auto-downgrading; commit-count and LOC are starting signals, not findings.
- V6. Implementation vs proxy: a verified proxy with an unverified implementation is effectively unverified. State whether the implementation is verified separately.

EVIDENCE DISCIPLINE (read before writing findings[]):
- Do not assert a specific deploy-commit SHA, bytecode equivalence, or "identical to audited commit" unless you actually fetched the artifact that shows it (e.g., a deployed-addresses JSON you opened, an explorer page you read). Inferred or plausible matches belong in unknowns[], never in findings[] or evidence[].
- Evidence[] entries must correspond to pages/files you actually retrieved this run. A URL you did not open is not evidence.

Then write the steel-man section per Hard Rule 11.

Grade rules:
- green   = deployed bytecode verified on the explorer (proxy AND implementation if proxied; "Similar Match" on a standard proxy pattern is fine per V1), a public source repo exists whose contents correspond to the explorer-visible source, AND ≥1 audit from a recognized firm covering the currently-deployed contracts (≤6 months of drift OR drift was re-audited). A missing local compile-match is not a downgrade — record it in unknowns[] and still grade green if the other conditions hold.
- orange  = verified but with visible drift from the public repo, OR audit scope is stale relative to deployment, OR only minor / unknown-firm audits exist, OR only some of the main contracts are verified, OR proxy verified but implementation only partially verified.
- red     = unverified bytecode (or verified proxy with unverified implementation), OR no audit in protocol.audit_links, OR no public repo.
- unknown = reserved for when the protocol's verifiability posture genuinely cannot be assessed (e.g., explorer and repo both inaccessible for this protocol). Do NOT use unknown merely because you, the analyst, could not run a particular check such as a bytecode diff — that goes in unknowns[] while the grade is still assigned from the evidence you do have.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "verifiability",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

4 dimensions not yet assessed (Control, Ability to exit, Autonomy, Open Access)

Control unknown Unverified

Not yet assessed

No model has graded this dimension yet. Run the slice prompt through any LLM and submit the JSON — once ≥3 independent runs agree, the quorum bot merges the verdict here.

Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-cow-amm
- protocol.name:              Balancer CoW AMM
- protocol.chains:            xDai, Ethereum, Arbitrum, Base
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: CONTROL

Evaluate who can change the protocol's rules, how fast, and how broadly.

(Step 0 capability probe and the off-chain-only fallback live in the preamble — those rules apply here.)

### MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[])

- **C1.** For each address you assess: who is the contract owner / admin / pendingAdmin / governor — read these via the block explorer's "Read Contract" tab OR `https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=owner` (BARE method names: `&method=owner`, `&method=admin`, `&method=pendingOwner`, `&method=governor`). For Safes use `/api/safe/owners?chainId=<id>&address=0x...`. When a protocol has multiple major versions deployed (v2/v3/v4), perform C1 reads on the NEWEST deployment separately — newer deployments often have weaker control surfaces than the legacy core.
- **C2.** Upgrade mechanism: transparent proxy / UUPS / Beacon / Diamond / immutable. Identify the proxy admin address. Check upgradeability of GOVERNANCE contracts too — a Governor / Aragon Voting / OZ Governor is often itself a proxy whose admin is the Timelock. Asymmetry: when fund-holding cores are immutable AND governance has no admin path that reaches them, an upgradable Governor/Timelock is T3-only and must NOT drag the verdict below green on that basis alone (see grade rules and the "immutable cores" caveat). Only call upgradability "mixed" if you can name a concrete function on the upgradable surface that reaches T1 or T2 on user funds.
- **C3.** EXECUTION PATH (enumerate every stage, in order, with delays in seconds). The operative path is usually a chain (voting → scheduler → timelock → executor; or governor → queue → execute; or Aragon Voting → DualGovernance → EmergencyProtectedTimelock → AdminExecutor). For each stage, record (a) the contract address, (b) the delay constant name + value in seconds, (c) the URL you read it from (block-explorer Read Contract OR `/api/contract/read?...&method=MIN_DELAY&block=<n>`). Do NOT stop at the first timelock-shaped contract — if its admin is itself called by another contract, keep walking. The grading delay is the SUM OF DELAYS ON THE UNCONTESTED FAST PATH (shortest time a proposal with no opposition can go from submission to executable). Dynamic / contested extensions (veto signaling, rage quit, escrow delay) are modifiers, not the basis — note them separately.
- **C4.** Enumerate EVERY multisig with reachable control — main proxy admin, emergency activation, emergency execution, reseal / pause, gate-seal committees, tiebreaker, per-module admins. For each Safe, fetch threshold + owners + version via `/api/safe/owners?chainId=<id>&address=0x...` (response includes raw eth_call data, so the URL is citable evidence). Enumerate ops/council/incentives multisigs even when off the upgrade path — record their scope so a reader can see they are NOT on the upgrade path. For each: (a) address, (b) threshold / total signers, (c) signer identities classified as insider (team, paid auditors under ongoing engagement, mandated service providers) vs non-insider (independent community members, unaffiliated researchers), (d) the specific power held (upgrade, pause, parameter, etc.).
- **C5.** On-chain governance: Governor / GovernorBravo / OZ Governor / Aragon Voting with token-weighted voting? Record proposal threshold, voting period, quorum, and the timelock delay between queue and execute. Every numeric constant must come from a Read Contract call you can link to, or be in unknowns[] with the C-code. If votingDelay / votingPeriod are denominated in BLOCKS, convert to seconds at the chain's CURRENT block time (Ethereum mainnet ≈ 12s post-Merge, not the 15s in older Compound/Bravo deployments) — cite both block count and converted seconds.
- **C6.** EMERGENCY POWERS: separate emergency-pause / guardian role with a different time cap or different actor than the main upgrade authority? Record it explicitly.
- **C7.** POWER TIER (blast radius). For each privileged path in C3–C6, classify the WORST thing that path can do, choosing the highest applicable tier. Cite the specific function name and any on-chain bound — tier claims without a named function are unsupported.
    - **T1 — FUND-CRITICAL**: replace implementation of contracts holding user funds; change AMM math / accounting / collateral logic; mint unbacked debt or shares; pause withdrawals; drain user-fund treasury; change oracle to attacker-controlled source; replace upgrade admin with EOA.
    - **T2 — ECONOMICALLY MATERIAL**: change fee parameters within bounded ranges; redirect protocol fees; add/remove markets / collateral types; bounded inflation or token mint within hard-capped schedule; spend protocol-owned (non-user) treasury.
    - **T3 — GOVERNANCE-INTERNAL**: change voting rules, quorum, voting period, proposal threshold; upgrade the Governor itself; rotate Timelock admin; mint governance tokens within a capped annual schedule.
    - **T4 — OPERATIONAL**: incentives distribution, grants, ENS / frontend canonicalization, deployment coordination, periphery router deprecation.
  The grade is set by the HIGHEST tier reachable on the uncontested fast path, not the median. State the tier and the binding function in the verdict.

### Read Contract discipline (applies to C3, C4, C5)

Every numeric constant cited (timelock delays, voting periods, multisig thresholds, quorum percentages) must come from EITHER (a) a block-explorer Read Contract URL, OR (b) a DeFiPunkd `/api/contract/read` or `/api/safe/owners` URL (preferred with `&block=<n>` for content-addressed evidence), OR (c) an unknowns[] entry with the C-code. Docs / blog posts are corroboration only — they cannot be the sole citation for a value that is also readable on-chain.

### Off-chain-only substitute hierarchy (when grading_basis="off-chain-only" — see preamble Rule 16)

When on-chain reads were genuinely unreachable this run, eligible off-chain substitutes in priority order:
1. Linked audit PDFs (admin roles, multisig members, timelock delays usually enumerated).
2. Governance forum posts that quote constants from a successful on-chain proposal (cite post URL + linked execution-tx URL).
3. Official protocol docs pages with named addresses and roles (must be on a domain owned by the protocol).
4. GitHub README / SECURITY.md / governance/*.md at a pinned commit SHA.

Forbidden substitutes: third-party blog posts, X / Twitter threads, search-result snippets, model memory. Required degradation: any C-code citing a numeric constant from docs/forum/audit prose ONLY must also carry an `unknowns[]` entry with `-offchain` suffix noting "value not re-read on-chain in this run; corroboration only".

### Grade rules (apply the timelock bar conditional on the highest C7 tier reachable on the fast path)

Security Council standard (used below): a multisig qualifies as "Security Council" only if ALL of: ≥7 signers, ≥51% threshold, ≥50% non-insider signers, every signer publicly announced. Failing any criterion = NOT a Security Council, regardless of signer reputation.

- **green**: highest reachable tier is T3 or T4 regardless of timelock; OR T2 reachable with uncontested-fast-path delay ≥7 days; OR T1 reachable only via immutable contracts (T1 is unreachable); OR T1 reachable with uncontested-fast-path delay ≥7 days combined with a Security Council multisig; OR T1 reachable with uncontested-fast-path delay ≥7 days through active on-chain governance with broad token distribution.
- **orange**: T2 reachable with uncontested-fast-path delay >0 but <7 days; OR T1 reachable with uncontested-fast-path delay >0 but <7 days; OR a multisig failing one or more Security Council criteria sits on a T1/T2 path; OR unclear upgrade authority on a T1/T2 path; OR governance with very short timelock or low quorum on a T1/T2 path.
- **red**: T1 reachable with no timelock by a single EOA or 2-of-3 multisig; OR a T1 upgrade admin that is not a smart contract you can audit.
- **unknown**: completed the checklist but still cannot determine the upgrade authority OR cannot classify the highest tier reachable on the main contracts.

Tiering caveats:
- "Bounded" must be enforced ON-CHAIN to count as T2. A function that sets fees with no upper-bound check is T1 — cite the bound check.
- Recurring T2 economic extraction (e.g. fee redirect with no rate limit) approaches T1 over time. A single proposal that can permanently redirect all future revenue is T1.
- T3 assumes the governance contract cannot itself authorize a T1/T2 action without going through the same timelock. If governance can self-upgrade to bypass the timelock, T3 collapses into T1.
- Do not downgrade tier by hand-waving ("realistically governance would never…"). Tier on what the contract permits, not what feels likely.

Notes:
- **Dynamic / dual-governance timelocks** (Lido, Compound escrow veto): the rubric grades on the uncontested path because that is the path most upgrades take. A dynamic extension that fires only under stake-weighted opposition is a real protection — name it in the green steel-man, but it does not lift an orange fast path into green; state the tension in the verdict.
- **Immutable cores with upgradable governance** (Uniswap-style): if fund-holding contracts are immutable and have no admin-reachable function moving / freezing / re-routing user funds, the highest reachable tier on the upgrade path is T3 — green regardless of timelock. Don't grade this orange just because the Governor is a proxy — that's a C2 fact, not a downgrade. Downgrade only applies if you can cite a concrete function on the upgradable surface that reaches T1 or T2 (privileged hook, upgradable factory controlling fund-routing, fee-switch redirecting protocol revenue without bound).
- **The 7-day bar** reflects the exit-window standard — users need notice after a queued upgrade to withdraw if they disagree. The ability-to-exit slice grades the exit side; this slice grades the delay side; both must hold for users to actually benefit from the delay.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "control",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Ability to exit unknown Unverified

Not yet assessed

Whether users can exit on their own terms if the team disappears or acts adversarially. Requires per-protocol review; not available at Phase 0.

No model has graded this dimension yet. Run the slice prompt through any LLM and submit the JSON — once ≥3 independent runs agree, the quorum bot merges the verdict here.

Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-cow-amm
- protocol.name:              Balancer CoW AMM
- protocol.chains:            xDai, Ethereum, Arbitrum, Base
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: ABILITY-TO-EXIT

Evaluate whether users can withdraw their funds on their own terms, even under adversarial admin conditions.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- E1. Enumerate every user-facing exit function in the main contracts: withdraw, redeem, burn, requestWithdrawal, claim, exit, etc. List them by name. Do NOT treat the contract as a monolith.
- E2. For EACH exit function in E1: identify its access modifiers and any pause guards (e.g. _checkResumed, whenNotPaused, onlyRole). Functions that gate REQUEST PLACEMENT often differ from functions that CLAIM FINALIZED FUNDS — check both separately.
- E3. For each pause guard: identify the role holder (which address holds PAUSE_ROLE / GUARDIAN / etc.) and the maximum pause duration. Specifically check whether PAUSE_INFINITELY (or equivalent uncapped pause) is callable, and which actor can call it (single multisig vs governance vote). For role-holder reads use https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=hasRole&args=0x...,0x... or &method=getRoleAdmin&args=0x.... For "is currently paused" checks use &method=paused or &method=isPaused&args=<resume-code>. Use the BARE method name (no parens). Cite the URL with &block=<n> in evidence[].
- E4. EMERGENCY vs GOVERNANCE pause distinction: many protocols have a fast-acting emergency pause capped at N days and a slow governance pause that can be indefinite. Record both paths separately if present, with their time caps and actor classes.
- E5. Queued redemption: documented maximum queue duration, daily withdrawal caps, whether the queue itself is pausable.
- E6. Forced-exit / escape-hatch / permissionless emergency-exit mechanism for adversarial-admin scenarios.
- E7. Frontend dependency: confirm exit functions are directly callable on-chain (e.g. via Etherscan write tab or a generic wallet) without the project's frontend.

Then write the steel-man section per Hard Rule 11. Common red-vs-orange tension on this slice: indefinite pause exists (suggests red) BUT the realistic emergency path is time-capped AND claims of already-finalized exits are not pause-gated (suggests orange). Resolve this by stating who can do what for how long, not by stopping at the worst-case sentence.

Grade rules:
- green   = permissionless exit; pause is either absent, narrowly scoped to clearly-described emergencies with auto-expiry, or capped at ≤7 days; no frontend dependency for exit; claims of already-finalized exits are not pause-gated under any path.
- orange  = pausable with broad scope OR indefinite pause is reachable only through governance vote (not unilateral admin action), OR queued redemption with documented max > 7 days, OR claims-of-finalized are exempt but new-request placement can be paused indefinitely by governance.
- red     = exit requires admin signature, OR ANY actor (including governance) can pause CLAIMS of finalized exits indefinitely, OR there is no on-chain exit function at all (purely custodial), OR pause is held by a single EOA / 2-of-3 multisig with no time cap.
- unknown = checklist incomplete after checking the sources above.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "ability-to-exit",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Autonomy unknown Unverified

No Phase-0 autonomy signal

No model has graded this dimension yet. Run the slice prompt through any LLM and submit the JSON — once ≥3 independent runs agree, the quorum bot merges the verdict here.

Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-cow-amm
- protocol.name:              Balancer CoW AMM
- protocol.chains:            xDai, Ethereum, Arbitrum, Base
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: AUTONOMY

Evaluate this protocol's autonomy: can a failure of anything outside its own contracts cause theft or loss of user principal, loss of unclaimed yield, or materially change the protocol's expected performance? "Autonomy" is not "has zero external touchpoints" — it is "external failures are either survivable or recoverable without user loss". This slice adapts DefiScan v1's Autonomy dimension; dependencies are one of several criteria, not the whole frame.

GUARDRAILS (read before grading):
- Category alone (Liquid Staking, Bridge, RWA Lending, Restaking, …) does NOT force a grade. A category is a hint about where to look; the grade must come from the concrete A1–A9 findings below.
- Base-chain consensus (Ethereum PoS, the chain's validator set, the canonical Deposit Contract at 0x00000000219ab540356cBB839Cbe05303d7705Fa) is the SUBSTRATE, not a dependency, for any protocol deployed on that chain. Do not list "depends on Ethereum" as a finding.
- Oracles or other integrations used by DOWNSTREAM protocols that happen to read this protocol's token (e.g. Chainlink stETH/USD consumed by Aave) are NOT this protocol's dependencies. Count only what THIS protocol's contracts call or trust on-chain. If a feed or contract appears in the protocol's docs only as reference material for third-party integrators, EXCLUDE IT ENTIRELY — do not log it as a finding even with a "peripheral" or "referenced only" caveat.
- "Upgradeable admin can change things" belongs to the CONTROL slice; only count it here when the upgrade surface lets governance silently swap an external dependency (see A9).
- Underlying-asset risk in opt-in, isolated markets is NOT autonomy-red on its own. When a protocol wraps third-party yield-bearing assets (LSTs, LRTs, ERC-4626 vaults, lending receipts, restaked tokens) into per-market silos that users explicitly choose, a failure of one underlying does NOT propagate to other markets and is risk the user opted into per-market. Record it under A4 with depth and propagation scope, but do not let it alone drive a red verdict — red requires an external dependency that cross-cuts the protocol or that the user did not opt into at deposit time. A failure mode that is "if the LRT you deposited is hacked, your principal in that LRT-backed market is impaired" is the underlying's autonomy story, not this protocol's; grade it on whether THIS protocol introduces additional dependencies on top.
- Sub-module enumeration is mandatory before grading. If the protocol ships distinct product lines or modules (e.g., a v2 core AMM plus a newer perps/funding-rate module, a lending pool plus a separate vault layer, an L1 core plus a cross-chain extension), enumerate each in findings and grade against the WORST module weighted by its share of TVS or its blast radius. A green core does not rescue an orange/red sub-module; conversely, a small red sub-module with capped TVS may bound the overall grade to orange. Name each module by its on-chain factory or router address. If you do not know whether a module exists, that is an unknowns[] entry, not silence. EXPLICIT TVS WEIGHTING: in the verdict, state each module's approximate share of total protocol TVS (use "~X%" if exact figures unavailable; check DeFiLlama or block-explorer balances on the module's main contract) and how that share informs the weighted grade. Format: "Module A holds ~X% of TVS (grade: <g>); Module B holds ~Y% (grade: <g>); weighted overall = <grade> because <reason>." If a red sub-module holds <5% of total TVS and is capped, the overall grade may be orange; if it holds >25%, the overall grade is red. Do not let qualitative reasoning substitute for the percentages — write the numbers.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. External contract calls. Enumerate every external contract the core contracts call or read from (oracles, price feeds, AMM pools, lending pools, staking/deposit contracts, yield wrappers). For each, identify the address, the provider, and what user-facing function of this protocol would break or mis-price if that external contract paused, mis-reported, or behaved adversarially. Grep for "oracle", "aggregator", "getPrice", "latestAnswer", "chainlink", "pyth", "redstone" as a starting point. To verify an oracle is live and what it currently reports, hit https://defipunkd.com/api/contract/read?chainId=<id>&address=<oracle>&method=latestAnswer (or &method=latestRoundData, &method=getPrice, &method=description; use the BARE method name without parens) and cite the URL — the response includes blockNumber/blockHash and rawReturnData, which is stronger evidence than a docs page about which feed "should" be used.
- A2. Off-chain actor committees reporting INTO the protocol. Oracle committees, guardian multisigs, DAO-selected validator sets acting as protocol reporters, exit-bus signers, fraud-proof challengers. For each, record committee size, quorum, who picks the members, and what mis-reporting could do (mint, burn, finalize withdrawal, freeze). Distinguish this from governance admins (control slice). NOTE ON STAKING PROTOCOLS: validator slashing and node-operator misbehavior are properties of the base-chain substrate, NOT external dependencies, when the operator set is diversified enough that a coordinated failure caps at <5% principal loss. Count validator/operator risk under A2 ONLY if the operator set is small, non-diversified, or lacks bonding / slashing-insurance / diversification mitigations. A curated set of 30+ independent operators with documented diversification falls under the mitigated path; a 3-operator LST does not. Do not cite the protocol's own risk disclosure as evidence that operator failure = principal loss unless you also check the diversification and bond mitigations.
- A3. Bridge / cross-chain messaging dependencies. Only count bridges that carry material TVL or are required for a core user flow. For each, name the bridge operator (canonical L1↔L2, LayerZero, Wormhole, Axelar, custom multisig), the trust model (canonical, optimistic, light-client, guardian set), and what fraction of TVL or users ride on it. "wstETH exists on 15 chains" is not a finding unless material TVL sits there. Before listing any non-primary chain deployment as a dependency, verify it is still operational as of analysis_date — retired or sunset deployments (e.g., Lido-on-Terra, Lido-on-Solana) belong in unknowns[] or should be omitted, never cited as a current dependency.
- A4. Nested collateral / restaking chains. For restaking / LRT / receipt-of-receipt designs: record the depth of the collateral chain, every actor with slashing or freezing power at each level, and whether a failure N levels deep propagates to user principal here.
- A5. Fork lineage (silent check). If DeFiLlama's forkedFrom is non-empty, record it as one finding and move on. If empty, do NOT add a placeholder finding; it adds noise.
- A6. Fallback mechanisms and circuit breakers. What catches an external failure? Sanity-check contracts on oracle reports, rebase bounds, pause paths triggered by bad prices, second-opinion oracles, max-per-block throughput caps, withdrawal queues that absorb bad reports. Record which A1–A4 risks are mitigated by which fallback, and which are unmitigated. For EACH fallback you cite, state its activation status explicitly: (i) LIVE and enforcing on-chain today, (ii) DEPLOYED but not yet wired / activated (e.g., interface exists but the address is zero or the role is unassigned), or (iii) DOCUMENTED / PROPOSED only (forum post, LIP draft, audit pending). Only (i) counts as mitigation for the grade. A fallback in state (ii) or (iii) should be noted but must not reduce the risk in your steel-man or verdict. If you cannot determine activation status, add an unknowns[] entry rather than assume it is live.
- A7. Sequencer / L1-liveness dependency BEYOND the base-chain substrate. SCOPE — sequencer risk only counts here when the protocol IS its own L2/L3 appchain or app-rollup, where the sequencer is part of the protocol's own stack and a freeze is a protocol-level outage. A protocol permissionlessly deployed on a third-party L2 inherits that L2's sequencer as substrate, not an A7 dependency. Record the sequencer/DA trust model when A7 applies.
- A8. Keeper / relayer / off-chain bot liveness. Protocols that need permissionless-but-necessary off-chain actors (liquidation bots, auto-compounders, deposit relayers, intent solvers). Record whether the role is permissionless, what degrades if nobody runs it (yield paused, bad debt accumulates, positions go stale), and whether the failure mode is graceful or catastrophic.
- A9. Governance-mutable dependency surface. Can an admin or DAO action silently INTRODUCE a NEW EXTERNAL dependency — swap the oracle address to a different provider, register a new staking module that calls an untrusted contract, add a new bridge, route SY through a new external vault — without an exit window for users? Check the upgrade / router / module-registry contracts. Answer: which EXTERNAL dependencies are governance-mutable, who holds that power, and whether there is a timelock or exit window. SCOPE LIMIT — read carefully: A9 is about the *external dependency surface*, not the upgrade surface in general. "The proxyAdmin / EOA can upgrade the router implementation to arbitrary bytecode" is a CONTROL-slice finding (admin can rug), NOT an autonomy-A9 finding. A9 fires only when the upgrade specifically swaps out or adds a contract that THIS protocol calls or trusts (e.g., changing the oracle address from Chainlink to a malicious feed, registering a new SY adapter that points to a third-party vault, redirecting a bridge endpoint). If the only finding is "admin can change implementation," do not log it under A9 and do not let it drive the autonomy grade — note it under control instead. The autonomy-relevant version of the same upgrade key is "admin can swap [specific external address X] without timelock"; that requires identifying the specific external dependency that becomes mutable.

STEEL-MAN (per Hard Rule 13): write one-sentence strongest arguments for red, orange, and green using the A1–A9 findings.

IMPACTED TVS ESTIMATE: the headline MUST include a rough impacted-TVS figure — the fraction of protocol TVS that could be lost or frozen if the worst-unmitigated dependency you identified failed. Use "~X%" if exact numbers are unavailable, "<1%" for de minimis, "unclear" only if A1–A9 left the question genuinely open (in which case grade=unknown is usually correct). Do NOT substitute qualitative phrases like "significant" — give a number or bracket.

GRADE ANCHORS (mapped to DefiScan v1 stages):
- green = Stage 2 equivalent. Failure of any external dependency cannot cause loss of user principal or unclaimed yield. Either there are no material external dependencies, or every critical one has a documented fallback (A6) that keeps users whole. Governance cannot silently introduce new dependencies without an exit window (A9). Impacted TVS under any single-dependency failure: effectively 0.
- orange = Stage 1 equivalent. Failure of some external dependency can cause loss of unclaimed yield, or can materially change expected performance (pause withdrawals, freeze positions, degrade price quality), but cannot cause loss of principal. Committee-based oracles with sanity checks, canonical-only bridges, fallback paths that exist but are incomplete, or governance-mutable dependencies protected by a ≥7-day timelock. Impacted TVS is bounded and recoverable.
- red = Stage 0 equivalent. Failure of an external dependency CAN cause theft or loss of principal. Examples: single-provider oracle with no sanity check or fallback, material TVL on a non-canonical bridge with a guardian multisig, governance can hot-swap oracles or add staking modules with no timelock or exit window, unmitigated keeper-liveness dependency where positions become insolvent if bots stop. Impacted TVS is material.
- unknown = checklist incomplete after inspecting source + verified contracts. Prefer unknown over guessing when A1/A6 could not be reconstructed. RESERVE unknown for cases where the CORE ARCHITECTURE itself is unverifiable — not for cases where you merely cannot enumerate every per-market dependency in a multi-market protocol. If the core router/factory/oracle architecture is verifiable on-chain and you can determine whether the core requires external dependencies, grade the architecture even when an exhaustive per-market external-dependency census is infeasible. Acknowledge the per-market gap in unknowns[] but still issue a grade. Refusing to grade a multi-market protocol because you cannot list every SY/vault/market is over-use of unknown; grade the architecture and say so.

PROMPT-META CHECK (per Hard Rule 17): before finalizing, verify the verdict cites concrete contract addresses, docs, or code — not the rubric itself. If your verdict says "the protocol belongs to a category the rubric marks red", rewrite it with the A1–A9 finding that actually justifies the grade, or drop to grade=unknown.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "autonomy",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Open Access unknown Unverified

Not yet assessed

No model has graded this dimension yet. Run the slice prompt through any LLM and submit the JSON — once ≥3 independent runs agree, the quorum bot merges the verdict here.

Submit run ↗

### Per-protocol context (ground truth for this run)

- protocol.slug:              balancer-cow-amm
- protocol.name:              Balancer CoW AMM
- protocol.chains:            xDai, Ethereum, Arbitrum, Base
- protocol.category:          Dexs
- protocol.website:           https://balancer.fi
- protocol.github:            balancer
- protocol.audit_links:       https://github.com/spearbit/portfolio/blob/master/pdfs/Balancer-Spearbit-Security-Review-September-2024.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2021-04-balancer-balancerv2-securityreview.pdf, https://github.com/trailofbits/publications/blob/master/reviews/2024-12-balancer-v3-securityreview.pdf
- snapshot.generated_at:      2026-06-01T11:27:13.878Z
- analysis_date:              2026-06-08
- prompt_version:             29
- address_book:               null

### Pre-built read-API surfacer URLs (verbatim — fetchable as-is)
(no addresses pinned in this run — discover candidates from fetched website / GitHub / audit / explorer pages, record discovered addresses in evidence[] and protocol_metadata.admin_addresses, and put any reads you couldn't perform in unknowns[]. The next assessment will inherit your discoveries.)


---

You are contributing a single-slice assessment to defipunkd, a git-native transparency registry for DeFi protocols. Your JSON output will be attached to a pull request and compared against ≥2 other independent runs. Disagreements are surfaced publicly — be conservative, cite everything, return grade="unknown" when a signal cannot be determined, and resist stopping at the first damning finding.

### Per-protocol context (do not infer; these are ground truth)

A "Per-protocol context" block (provided alongside this prompt — either as a user message in API mode, or as a section appended below in copy-paste mode) lists the pinned inputs (protocol.slug, name, chains, category, website, github, audit_links, snapshot.generated_at, analysis_date, prompt_version, address_book) and a list of pre-built read-API surfacer URLs for each pinned address. Treat those values as authoritative for this run.

The pre-built surfacer URLs are accepted by your fetch tool's allowlist because they appear verbatim in the per-protocol context. Each surfacer page pre-executes the contract's zero-arg view methods and renders any address-typed return values as /address/{chainId}/0x… links inline — those rendered links are also fetchable post-fetch. /api/contract/read and /api/safe/owners JSON responses include a top-level `crawl.surfacers` array of /address/{chainId}/0x… URLs for every address-typed value in the result, so you can crawl directly from API responses. For addresses surfaced from non-defipunkd sources (block-explorer pages, GitHub, audit PDFs), the allowlist will reject your generated surfacer URL — record the address in protocol_metadata.admin_addresses and add a checklist-coded entry to unknowns[]; the next run will inherit it as a fetchable surfacer.

### Step 0 — Capability probe (do this before producing JSON)

Identify what fetch tool your environment exposes — `web_fetch`, `browser`, `url_context`, `google_search` (with underscore, colon, or space), Bing grounding, Perplexity, `web.run`, or anything similar. Search-grounding tools ARE valid fetch paths and you should use them; they are not forbidden. The distinction that matters is what the tool returned, not its name: a grounded response with the underlying page body is sufficient evidence; a bare 1–3-line preview snippet is not.

Probe: attempt one fetch of `protocol.website` and one fetch of a pre-built surfacer URL above (or the protocol's primary block-explorer page if no surfacer is pinned). Record the tool you used, the URL, and what came back (response body, HTTP status, allowlist rejection text, or "tool not present: <name>"). If either probe succeeded, proceed.

If both direct probes failed, **DO NOT STOP** — try search. Issue at least two broad search queries (protocol name + "Etherscan" / "official docs" / "GitHub" / "audit"; `site:` operators against etherscan.io / github.com / the docs domain). If a search query returns the underlying page body, that page URL is fetched evidence and goes in evidence[]. Set `grading_basis: "off-chain-only"` and continue with whatever you can extract. Empty `evidence[]` + `grade="unknown"` is only valid if BOTH direct probes AND ≥2 search queries returned nothing usable; in that case, log the attempted queries and observed-zero outcomes in unknowns[]. A submission claiming "blocked" without those receipts is treated as a non-submission (zero quorum weight). Search engines do not index raw JSON API endpoints — zero results for `defipunkd.com/api/contract/read` URLs is expected and is itself a recordable failure mode, not a system restriction.

### Anti-fabrication (the most important rule on this page)

Memory is not evidence. Treat your training data, prior conversations, and general knowledge as suggestions for what to look up — never as citations. Every URL in evidence[] must have been fetched in this run via any tool exposed in your environment, OR pasted into this conversation by the user. Constructing a URL is fine, but every variable part (address, commit SHA, repo path, contract name, method, args, block number) must come from a fetched/pasted source in this run, and the URL must then have been successfully fetched before it appears in evidence[]. URLs constructed from remembered addresses, repo paths, contract names, or guessed API methods are fabrication.

Before emitting JSON, run an evidence ledger check on every evidence[] entry:
1. The exact URL appears in your fetch transcript or in a user-pasted source body.
2. The fetched/pasted body contains the fact you're citing in evidence[].shows.
3. Every rationale.findings / protocol_metadata claim that depends on this evidence follows directly from that body, without recourse to memory.
4. Derived (rather than verbatim) claims are explicitly labelled as derived in evidence[].shows.
5. `fetched_at` is set ONLY when you actually fetched the URL in this run; if no timestamp is available, omit the field — never invent one.

If any check fails, remove the evidence entry and demote dependent claims to unknowns[]. Set grade="unknown" if demotion empties the grading basis. Do not ask the user to paste anything; do not withhold JSON; do not improvise from memory.

A plausible-sounding answer backed by unsupported evidence is WORSE than grade="unknown" — it pollutes the quorum. If the assessment requires leaning on remembered public facts ("Lido is governed by LDO token-weighted voting"), historical reports, common knowledge, or likely-architecture reasoning ("UUPS proxies typically have an admin role"), return grade="unknown" with specific unknowns[] entries. Optimize for reproducibility, not completeness — if a reviewer can't re-verify each claim from the evidence URLs alone, the claim does not belong in the JSON. Empty unknowns[] on a non-trivial protocol is a red flag, not a quality signal.

When the address_book is null/empty, you do not yet know any deployed address. Discover candidates from fetched website / GitHub / audit / explorer-search pages — addresses you "remember" from training data are not eligible, even for famous tokens (USDC, WBTC, stETH, UNI). If no address can be discovered after a good-faith attempt, return grade="unknown" with checklist-coded unknowns[] entries; do not invent addresses to fill the gap.

### Hard rules

1. Source classes that count as evidence:
   a) Public block explorers (etherscan.io, basescan.org, arbiscan.io, optimistic.etherscan.io, polygonscan.com, bscscan.com, snowtrace.io, scrollscan.com, lineascan.build, blastscan.io, era.zksync.network) for pinned addresses or addresses you discover transitively from them.
   b) The linked GitHub repos, at a specific commit SHA recorded in evidence[].commit.
   c) The audit PDFs / reports linked above.
   d) DeFiLlama's pinned fields (for category / chain lists only — not for risk assessment).
   e) DeFiPunkd's read API at https://defipunkd.com/api/{contract,safe}/... — see "On-chain reading" below.
2. If a signal cannot be determined after checking these, set grade="unknown" with ≥1 entry in unknowns[] naming what you looked for.
3. Every factual claim in rationale must map to ≥1 evidence[] entry.
4. Output exactly one JSON object matching the contract at the end, wrapped in a single ```json fenced code block. This rule applies to your FINAL assistant message only — issue tool calls freely during the run; tool-call reasoning lives in your model's tool-use channel and is not subject to this rule. Nothing before or after the fence — no prose, no follow-up questions, no requests for the user to paste anything. If evidence is incomplete, the correct response is still JSON, with claims demoted to unknowns[] and grade="unknown".

### Format rules (validation will reject submissions that violate these)
5. evidence[].url must be a bare `https://...` string — NEVER markdown link syntax. WRONG: `"url": "[Etherscan](https://etherscan.io/...)"`. RIGHT: `"url": "https://etherscan.io/..."`.
6. evidence[].commit, when present, must match `^[0-9a-f]{7,40}$` (lowercase hex, 7–40 chars). NEVER branch names or tags. Omit if you cannot pin a SHA.
7. evidence[].fetched_at, when present, must be ISO-8601 UTC (e.g. `2026-04-23T11:20:00Z`). Set whenever you actually fetched in this run.
8. evidence[].address, when present, must be `^0x[0-9a-fA-F]{40}$` (mixed case OK; checksum not validated).
9. Checklist codes (used in findings[].code and as unknowns[] prefixes) match `^[A-Z][A-Za-z0-9-]{0,16}$` — start with an uppercase letter, then digits / letters / hyphens. Examples: `E1`, `A3b`, `C2-emergency`, `V4-auditor`. No parens, spaces, dots, slashes, underscores. Use slice-defined codes verbatim. unknowns[] entries must be prefixed with the relevant code + colon (e.g. `"A3b: frontend fetch failed"`).
10. chat_url: ALWAYS null. Default share links (claude.ai/chat, chatgpt.com, gemini.google.com) require viewer login and are not publicly readable. The user enables "Share publicly" after you respond and pastes the public URL into the JSON before opening the PR.

### Thoroughness rules

11. Each per-slice body contains a "MANDATORY INSPECTION CHECKLIST". Every item must EITHER produce an evidence[] entry OR a specific unknowns[] entry naming it by code. Silent skips are rejected as incomplete.
12. Before assigning a grade other than "unknown", rationale.steelman must contain a one-sentence strongest argument for each of red / orange / green, and rationale.verdict must state which fits the evidence and why. If the steel-man for the chosen grade is weaker than for an adjacent grade, you have probably picked the wrong grade. When grade="unknown", set steelman to null and use verdict to summarize what blocked the assessment.
13. Distinguish actor classes (EOA, 2-of-3 multisig, 4-of-7+ multisig with identified signers, emergency-scoped time-capped multisig, on-chain governance vote with timelock) and function classes (claim-of-finalized vs new-request-placement, deposit vs borrow, mint vs redeem) — say which actor holds which power, on what time bound. "An admin can pause" is insufficient.
14. For on-chain slices (control, ability-to-exit, autonomy, verifiability), evidence[] must include ≥1 on-chain URL: a block-explorer URL OR a DeFiPunkd /api/{contract/read,safe/owners} URL (preferred — content-addressed when block-pinned). /api/contract/abi alone is metadata, not on-chain evidence. Source repos tell you what code SHOULD do; deployed contracts tell you what it ACTUALLY does. The "open-access" slice is exempt when claims are entirely about frontend / off-chain operator behavior. CARVE-OUT: `grading_basis="off-chain-only"` is exempt from this rule but is downweighted by the quorum bot regardless of grade.
15. Prompt-meta-check: if your verdict quotes prompt language as evidence ("the protocol meets the 'documented fallback' condition"), re-do the verdict — the prompt describes the rubric, not the protocol. Evidence cites what THIS protocol does, not what the rubric says protocols of type X do.
16. `grading_basis` is one of `"on-chain"` (default; omit field), `"off-chain-only"`, or `"mixed"`. Describes what was READ this run, not the verdict — `grade="unknown"` is allowed under any basis. Set `"off-chain-only"` ONLY when BOTH (a) ≥1 successful fetch of a docs / forum / audit / GitHub URL appears in evidence[], AND (b) ≥1 failed on-chain fetch attempt for THIS run is recorded in unknowns[] with a `-offchain` suffix. Empty `evidence[]` with `"off-chain-only"` is a category error and is rejected by the validator. Set `"mixed"` when some checklist codes were on-chain and others fell back. ABI-only finds (you read the ABI but couldn't read live state) are valid as `grade="unknown"` with the ABI cited; do not infer a grade from ABI shape alone.

### On-chain reading via the DeFiPunkd API

Don't encode calldata, decode return data, or guess at ABIs by hand. Three deterministic GET endpoints return JSON with blockNumber, blockHash, raw calldata, and rawReturnData — content-addressed when block-pinned (`&block=<n>`):

  ABI (auto-resolves proxies):
    https://defipunkd.com/api/contract/abi?chainId=<id>&address=0x...
  View call (any view method on the merged ABI; flat scalar args):
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=getOwners
    https://defipunkd.com/api/contract/read?chainId=<id>&address=0x...&method=balanceOf&args=0x...
  Safe (threshold + owners + version in one call):
    https://defipunkd.com/api/safe/owners?chainId=<id>&address=0x...

Use the BARE method name in `&method=` (e.g. `&method=totalSupply`, NOT `&method=totalSupply()`). Browser tools normalize `(` to `%28` and reject the normalized URL; bare names dodge that. Pass arguments via `&args=` (comma-separated, declaration order). Append `&block=<n>` for content-addressed evidence.

Supported chainIds: 1 (ethereum), 10 (optimism), 56 (bsc), 130 (unichain), 137 (polygon), 324 (zksync), 8453 (base), 42161 (arbitrum), 43114 (avalanche), 59144 (linea), 81457 (blast), 534352 (scroll), 11155111 (sepolia).

Use this API for any factual claim about contract ABI shape, view-method return values (owner(), getOwners(), getThreshold(), totalSupply(), implementation(), paused(), MIN_DELAY(), …), and Safe membership. Do NOT invent ABIs or return values from training data.

Note on noisy address_book: the pinned address_book is sourced from prior assessments. It may include token deployments, oracle feeds, peripheral contracts, or mis-classified entries. Skip surfacer URLs that don't fit your slice — fetch only those whose role hints suggest control / pause / upgrade authority.

### Protocol metadata refresh (populate `protocol_metadata` in the output)

DeFiLlama's pinned inputs may be stale or wrong. As a side-effect of this assessment, populate `protocol_metadata` with anything you verify. Leave fields null / empty arrays if you did not verify them — do NOT echo pinned inputs through; null means "not re-verified this run".

- `github`: array of canonical source-code repo URLs.
- `docs_url`: canonical developer / protocol documentation site.
- `audits`: array of `{ firm, url, date }` (date as YYYY-MM or YYYY-MM-DD).
- `governance_forum`: primary discussion forum URL (Discourse, Commonwealth, etc.).
- `voting_token`: `{ chain, address, symbol }` or null. Omit if not token-governed.
- `bug_bounty_url`: public bug bounty page (Immunefi, HackerOne, self-hosted).
- `security_contact`: private-disclosure channel — security@ email or SECURITY.md URL. Distinct from public bug bounty.
- `deployed_contracts_doc`: docs page that lists deployed addresses per chain. Don't enumerate; just link.
- `admin_addresses`: array of `{ chain, address, role, actor_class }` for multisig / timelock / owner / proxy-admin addresses. `actor_class` ∈ `"eoa" | "multisig" | "timelock" | "governance" | "unknown"`. These are the anchors for future runs' address_book.
- `upgradeability`: `"immutable" | "upgradeable" | "mixed" | "unknown"`. "mixed" = some core contracts immutable, others behind proxies.
- `about`: 2–4 sentence plain-English description. Name the user action (stake, borrow, swap, bridge, mint, redeem), the asset / market, and the distinctive mechanism (liquid staking receipt, isolated lending pools, constant-product AMM, intents auction, etc.). Do not restate category / chains / TVL.

Every non-null field in `protocol_metadata` must be backed by ≥1 entry in evidence[].

---

### Slice: OPEN-ACCESS

Evaluate who is allowed to use the protocol and whether any of that permission is granted off-chain.

Scope: this slice is about ADMISSION — who can enter, exit, or transact. Operator LIVENESS (what breaks if keepers/oracles go offline) is assessed in the dependencies slice and is out of scope for the grade here. You may note operator dependencies as context, but do not let "the protocol halts if operator X disappears" drive the access grade on its own; that belongs in dependencies. Source verification / contract verification on block explorers is assessed in the verifiability slice and is out of scope here — do NOT let "contract is unverified" drive the access grade.

Framing: the smart contracts are the access layer; frontends are UX. A permissionless contract is reachable by any client (SDK, third-party UI, aggregator, wallet integration). Frontend ToS, IP geo-blocking, and wallet screening are publisher policies on one specific client — they are reported as context but do NOT determine the grade. The grade hinges on (1) what the contract itself permits, and (2) whether the protocol is practically reachable without the official publisher's cooperation.

Meta-check before finalizing: if your verdict cites phrases from this prompt as evidence ("the protocol meets the 'credible alternatives' condition", "this fits the 'documented fallback' rule"), redo the verdict. The prompt describes the rubric; evidence must come from the protocol. A verdict should cite what the protocol does, not what the rubric says.

MANDATORY INSPECTION CHECKLIST (every item below must appear in evidence[] OR unknowns[]):
- A1. Whitelist / allowlist modifiers in user-facing entry points. Grep for "onlyWhitelisted", "onlyRole", "allowlist", "isAccredited", "isKYCed". Note which functions are gated and who can add/remove from the list.
- A2. Off-chain operators in the admission path: keepers, sequencers, privileged relayers, oracle posters whose approval is required to admit a user action (not just to keep the protocol live). For each, identify whether the role is held by a single operator, a permissioned committee, or is permissionless. Enumerate per user-facing function class (deposit vs withdraw-request vs claim-finalized vs transfer) which ones require operator approval to be admitted, and which ones admit users unconditionally. A function whose placement is unconditional but whose downstream settlement depends on an operator is an admission-permissionless function — flag the liveness dependency as context and defer its grading weight to the dependencies slice.
- A3. Frontend restrictions on the official interface — record as context, not as a grade lever. Distinguish:
    - A3-passive: boilerplate ToS clauses (sanctions attestation, restricted-territory self-certification, VPN-circumvention prohibition, "comply with applicable law" eligibility, age of majority).
    - A3-active: runtime enforcement — IP-based geo-blocking, wallet-address screening against a sanctions oracle (Chainalysis, TRM, Elliptic), KYC wall, rendering-blocking jurisdiction banner.
  Record findings under the correct tier. Quote ToS text or banner text in evidence[].shows. These findings populate the headline and rationale but do NOT move the grade by themselves; the grade is set by A1, A2, and the A3b path check below.
- A3b. Independent access paths (the operative grade input). Enumerate paths that do not require the official publisher's cooperation:
    - Published SDK / library / CLI for direct contract interaction.
    - Third-party frontends operated by separate legal entities.
    - Wallet-integrated access (MetaMask Swaps, Safe apps, etc.).
    - DEX / lending / yield aggregators that route through the contracts.
  Record at least one concrete link per path that exists. The protocol does NOT have to self-document these — the test is existence, not UX cost. An A3b-i redistribution of the official UI bound by the same ToS does NOT count as an independent path.
- A4. Sanctions / compliance tooling at the contract level: does the protocol check addresses against OFAC lists or similar on-chain blocklists in the contract itself? (Frontend-only screening belongs in A3.)
- A5. Differentiate read access vs write access: many protocols are read-permissionless (anyone can view state) but write-gated (only certain addresses can deposit/borrow). Record both.
- A6. ToS / Legal links: locate them on the website and produce a VERBATIM quote of any jurisdictional, sanctions, or eligibility clause in evidence[].shows. If you cannot extract the clause text verbatim (SPA render failure, paywall, dead link, etc.), do NOT paraphrase or infer from general knowledge — record the ToS URL in unknowns[] with the reason extraction failed. Assertions about ToS content without a verbatim quote will be downweighted by reviewers.

Then write the steel-man section per Hard Rule 11.

Grade rules (admission-focused; liveness concerns belong in dependencies; source verification belongs in verifiability):
- green   = no contract-level whitelist/KYC on user entry/exit; no operator approval required to admit a user action; AND at least one independent A3b path exists (published SDK, third-party frontend, wallet integration, or aggregator routing). Frontend ToS posture and A3-active enforcement on the official UI do NOT block green when contracts are permissionless and an independent path exists — they are reported as context.
- orange  = contracts admit users unconditionally, BUT the protocol is operationally captured by the official publisher: no published SDK, no third-party frontend, no wallet integration, no aggregator routing. The contract is theoretically open but practically reachable only through the official UI. Also applies when admission requires approval from a permissioned committee that is governance-managed with a documented replacement procedure.
- red     = contract-level whitelist / KYC on user entry/exit, OR admission of a core user action requires approval from a single privileged operator or a small committee with no documented replacement procedure, OR enforces an on-chain blocklist updatable by a single party.
- unknown = checklist incomplete after checking the sources above.

Default-grade guidance: when contracts are fully permissionless AND any A3b independent path exists, the default grade is green regardless of frontend ToS or A3-active enforcement on the official UI. Frontend geo-blocking, sanctions-oracle wallet screening, and ToS sanctions clauses are publisher policies on one client and are reported in findings/headline as context, not as grade levers. To grade orange on operational-capture grounds, the auditor must affirmatively show that ALL independent paths are absent or also gated.

Guideline on committees: where admission depends on a multi-operator committee, the relevant axes are (a) set size, (b) whether replacement/rotation is governed on-chain, (c) whether the replacement procedure is publicly documented. A large set with on-chain governance replacement should not be graded as a single-party operator even if rotation is not instantaneous. A small set with informal replacement should be treated as a single-party operator.

---

### JSON output contract

Return exactly one JSON object inside a single ```json fenced block. Shape:

{
  "schema_version": 4,
  "slug": "<copy protocol.slug from the per-protocol context>",
  "slice": "open-access",
  "snapshot_generated_at": "<copy snapshot.generated_at from the per-protocol context>",
  "prompt_version": 29,
  "analysis_date": "<copy analysis_date from the per-protocol context>",
  "model": "<exact model name, e.g. claude-opus-4-7 / gpt-5-thinking / gemini-3-pro>",
  "chat_url": null,
  "grading_basis": "on-chain | off-chain-only | mixed (optional; omit for on-chain)",
  "grade": "green | orange | red | unknown",
  "headline": "<one-line summary>",
  "short_headline": "<≤6 words, ≤80 chars; omit if you can't fit>",
  "rationale": {
    "findings": [{ "code": "E1", "text": "<concrete, source-cited finding>" }],
    "steelman": { "red": "<one sentence>", "orange": "<one sentence>", "green": "<one sentence>" },
    "verdict": "Choosing <grade> because <reason ranking one steel-man above the others, citing specific evidence>."
  },
  "evidence": [{ "url": "https://...", "shows": "<what this URL demonstrates>", "chain": "...", "address": "0x...", "commit": "<hex SHA>", "fetched_at": "2026-04-23T11:20:00Z" }],
  "unknowns": ["E3: <thing you looked for but couldn't determine>"],
  "protocol_metadata": {
    "github": ["https://github.com/org/repo"],
    "docs_url": "https://docs.protocol.xyz",
    "audits": [{ "firm": "Trail of Bits", "url": "https://...report.pdf", "date": "2025-09" }],
    "governance_forum": "https://forum.protocol.xyz",
    "voting_token": { "chain": "Ethereum", "address": "0x...", "symbol": "XYZ" },
    "bug_bounty_url": "https://immunefi.com/bounty/protocol",
    "security_contact": "security@protocol.xyz",
    "deployed_contracts_doc": "https://docs.protocol.xyz/deployments",
    "admin_addresses": [{ "chain": "Ethereum", "address": "0x...", "role": "DAO treasury multisig", "actor_class": "multisig" }],
    "upgradeability": "immutable | upgradeable | mixed | unknown",
    "about": "<2–4 sentences>"
  }
}

Rules recap:
- grade="unknown" ⇒ steelman=null; unknowns[] ≥1; evidence[] may be empty.
- grade!="unknown" ⇒ steelman={red,orange,green}; evidence[] ≥1; verdict starts with "Choosing ".
- findings[].code matches the slice's checklist prefix verbatim (E1, C2-emergency, V4a, …); unknowns[] entries are checklist-coded ("E3: …").
- Wrap in a single ```json fence; nothing before or after. URLs are bare strings, never markdown links.

Stage

Preview of the Phase-3 maturity framework. DeFiPunk'd will adopt DeFiScan v2's stages verbatim; the section is rendered below in its intended shape so the structure is visible today.

Balancer CoW AMM has not yet been assessed under the DeFiScan v2 stage framework.

The walkaway test is the central criterion. Once stages land, protocols reach Stage 1 only if users can exit in the presence of malicious operators even when the emergency council disappears.

Scope of assessment

Stage 0 requirements pending

Stage 1 requirements pending

Stage 2 requirements pending

Learn more about DeFiScan v2 stages →

Protocol Info

Security

[defillama] [:]

Audits: 16 audits

Spearbit — 2024-09 — balancer-v3
Trail of Bits — 2024-12 — balancer-v3
Trail of Bits — 2021-04 — balancer-v3
github.com — 2023-05-08 — balancer-v2
github.com — 2022-10-25 — balancer-v2
github.com — 2022-09-23 — balancer-v2
github.com — 2022-09-02 — balancer-v2
github.com — 2022-05-27 — balancer-v2
github.com — 2022-05-27 — balancer-v2
github.com — 2021-10-09 — balancer-v2
github.com — 2021-10-08 — balancer-v2
github.com — 2021-04-22 — balancer-v2
github.com — 2021-04-02 — balancer-v2
github.com — 2021-03-15 — balancer-v2
github.com — balancer-v2
docs.balancer.fi — balancer-v1
Bug bounty: https://immunefi.com/bug-bounty/balancer/information/
Security contact: security@balancer.fi

Technical

[:]

Voting token: BAL Ethereum: 0xba100000625a3754423978a60c9317c58a424e3D
Deployed contracts: https://docs.balancer.fi/developer-reference/contracts/deployment-addresses/mainnet.html
Upgradeability: Unknown

Provenance

[defillama]

Review status: listed
Updated: 2026-06-01 11:27 UTC

Deployments

Risk analysis

Stage

Risk analysis

Stage

Risk analysis

Stage

Risk analysis

Stage

Protocol Info

Links

Security

Technical

Provenance