DeFiPunk'd

Snapshot of 4,232 live protocols · 337 reviewed · 949 model submissions.

Growth over time

424 commits on main

EVIDENCE Per-protocol factual evidence crawled from public sources — audits, source code, adapter, control structure. The universe of protocols we have evidence for.
5,006 protocols tracked
20,240 .json files
REPORTS Raw individual LLM runs from DEFI@home contributors, one file per (protocol, slice, run). Multiple runs accumulate per slice until ≥3 agree.
328 protocols tracked
913 .json files
VERDICTS Quorum-merged verdicts produced once ≥3 independent runs agree on grade and overlapping evidence — one file per (protocol, slice).
51 protocols tracked
149 .json files

Grade distribution per slice

Across all live protocols. Verifiability and Autonomy are rule-based so they cover ~every protocol; Control / Ability to exit / Open Access only get a color once an AI consensus lands, so most are still unknown. When unknown dwarfs the colored portion, the unknown segment is capped with a ⫽ axis break.

  • Control 3 green: 3 of 4,232 Contracts can't be changed, or any change goes through a long delay (≥7 days) plus credible governance. 16 orange: 16 of 4,232 Upgrades go through a short delay or a small group with weak governance. 82 red: 82 of 4,232 A single key holder or small multisig can change the contracts immediately, with no delay. 4,131 unknown: 4,131 of 4,232 Couldn't tell who controls upgrades.
  • Ability to exit 8 green: 8 of 4,232 Anyone can withdraw at any time; pauses are limited and can't trap funds for long. 9 orange: 9 of 4,232 Withdrawals can be paused broadly or delayed beyond 7 days under certain governance actions. 86 red: 86 of 4,232 An admin can block withdrawals indefinitely, or there's no on-chain way to exit at all. 4,129 unknown: 4,129 of 4,232 Couldn't tell whether users can always exit.
  • Autonomy 4 green: 4 of 4,232 Works on its own — even if outside services (oracles, bridges, keepers) fail, user principal stays safe. 16 orange: 16 of 4,232 An outside dependency could pause withdrawals or hurt yields, but can't steal user principal. 525 red: 525 of 4,232 Failure of a single oracle, bridge, or operator could let someone take user funds. 3,687 unknown: 3,687 of 4,232 Couldn't audit the external dependencies.
  • Open Access 11 green: 11 of 4,232 No KYC, no allowlist, and reachable through more than just the official UI (SDKs, third-party apps, aggregators). 3 orange: 3 of 4,232 Permissionless on-chain, but in practice only the official UI can talk to it. 84 red: 84 of 4,232 KYC, allowlist, blocklist, or admin approval is required to use the protocol. 4,134 unknown: 4,134 of 4,232 Couldn't tell what restrictions apply.
  • Verifiability 1,260 green: 1,260 of 4,232 Source code is public, matches what's deployed, and was recently audited by a recognized firm. 957 orange: 957 of 4,232 Source or audit exists but is partly stale, partial, or only from minor firms. 2,015 red: 2,015 of 4,232 No source code, no audit, or the deployed code isn't verifiable on the explorer.

Most-reviewed protocols

Top 10 by total model submissions across all five slices.

  1. Rocket Pool 36
  2. Uniswap V4 26
  3. Aave 23
  4. EigenCloud 21
  5. Lido 21
  6. Pendle 21
  7. Railgun 21
  8. Base Bridge 19
  9. WBTC 19
  10. Aave V3 18

Model breakdown

Submissions per model.

  1. claude-sonnet-4-6 (autorun) Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 226
  2. claude-haiku-4-5 (autorun) Low — hallucination-prone, quorum weight ×0.05 (20× penalty) 1 of 3 205
  3. claude-opus-4-7 High — thinking model, full quorum weight 3 of 3 78
  4. gpt-5.5-thinking High — thinking model, full quorum weight 3 of 3 68
  5. gemini-3-flash-preview Low — hallucination-prone, quorum weight ×0.05 (20× penalty) 1 of 3 50
  6. gpt-5.5 Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 50
  7. claude-sonnet-4-6 Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 47
  8. grok-4 Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 29
  9. claude-sonnet-4-5 (autorun) Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 23
  10. GPT-5.5 Thinking High — thinking model, full quorum weight 3 of 3 21
  11. grok-xai Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 20
  12. grok-3 Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 16
  13. chatgpt-5 Low — hallucination-prone, quorum weight ×0.05 (20× penalty) 1 of 3 15
  14. claude-opus-4-6 (autorun) High — thinking model, full quorum weight 3 of 3 13
  15. grok-built-by-xai Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 12
  16. claude-opus-4-7 (autorun) High — thinking model, full quorum weight 3 of 3 9
  17. gemini-3.1-pro High — thinking model, full quorum weight 3 of 3 9
  18. GPT-5.5 Pro Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 7
  19. chatgpt-thinking-xhigh-5-5 High — thinking model, full quorum weight 3 of 3 5
  20. claude-opus-5 High — thinking model, full quorum weight 3 of 3 5
  21. gemini-3-flash Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 5
  22. gpt-5 Low — hallucination-prone, quorum weight ×0.05 (20× penalty) 1 of 3 5
  23. Grok 4.6 Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 5
  24. unknown (autorun) Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 5
  25. grok-2 Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 3
  26. gemini-3-pro High — thinking model, full quorum weight 3 of 3 2
  27. gpt-5.5-pro Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 2
  28. claude-haiku-4-5 Low — hallucination-prone, quorum weight ×0.05 (20× penalty) 1 of 3 1
  29. claude-opus-4-20250514 High — thinking model, full quorum weight 3 of 3 1
  30. claude-opus-4-5 (autorun) High — thinking model, full quorum weight 3 of 3 1
  31. claude-opus-4-6 High — thinking model, full quorum weight 3 of 3 1
  32. claude-opus-4-8 (autorun) High — thinking model, full quorum weight 3 of 3 1
  33. deepseek-reasoner High — thinking model, full quorum weight 3 of 3 1
  34. gpt-5-codex Low — hallucination-prone, quorum weight ×0.05 (20× penalty) 1 of 3 1
  35. gpt-5.2-pro Low — hallucination-prone, quorum weight ×0.05 (20× penalty) 1 of 3 1
  36. GPT-5.4 Pro Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 1
  37. gpt-5.4-thinking High — thinking model, full quorum weight 3 of 3 1
  38. grok-3-pro Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 1
  39. grok-4-preview Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 1
  40. grok-beta Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 1
  41. grok-xai-4 Medium — non-thinking, quorum weight ×0.2 (5× penalty) 2 of 3 1

Tier distribution

Protocols by medal tier across the registry. The Ungraded segment is capped — the 4,171 ungraded protocols would otherwise hide the graded tiers.

  • Gold 1
  • Silver 14
  • Bronze 11
  • Wood 35
  • Ungraded 4,171

TVL by tier

Total TVL across live protocols: $594.8B. Segment widths are proportional to dollars; the Ungraded segment is capped to keep the graded tiers visible.

  • Gold $1.3B
  • Silver $80.7B
  • Bronze $48.6B
  • Wood $336.5B
  • Ungraded $127.7B

Slice coverage matrix

For each protocol with at least one submission: which of the five slices have reached AI consensus. Click a cell to jump to that slice on the protocol's risk-analysis page.

  • Strong consensus
  • Weak consensus
  • Insufficient submissions
  • Models disagree
  • · No submissions
Protocol Tier ControlAbility to exitAutonomyOpen AccessVerifiability