Candidate evidence is labelled as such
Until it is released, collected evidence appears as “Candidate evidence · not approved” and never counts as coverage in the Explorer or the wallet.
Evidence AgentsAvailability: In development
Evidence Agents help gather published audit reports and their sources. AI output is never shown as verified coverage: human-reviewed release and deterministic matching against deployed code decide what counts.
AI helps find and organize sources. It never decides coverage.
Anything an agent extracts stays a candidate until it passes the checks below and people release it.
FIG 1Evidence candidate, as produced
Evidence candidate, as produced
public/evidence/demos/candidates/aave-v4-chainsecurity.json
{"schemaVersion": "poa-evidence-candidate-v1","origin": {"publisher": "ChainSecurity","fetchedAt": "2026-11-13T04:42:08Z","contentSha256": "2caa833f0b2b9ddc…"},"reviewedCode": {"reviewedCommit": "cdacec509e4f848b…","deploymentAddress": "0xabd0e26fe17bde4f1f1187ed8aa80c274e03d8b5"},"claims": [{"kind": "reviewed_revision","statement": "The report lists commit cdacec50 among the reviewed code versions.","locator": "page:6;section:Assessment Overview","evidenceLevel": "ai_extracted"},{"kind": "included_contract","statement": "The report lists SpokeInstance.sol in scope.","locator": "page:5;section:Scope","evidenceLevel": "ai_extracted"},],"codeLinkage": {"status": "deployed_code_matched","matchQuality": "exact_match"},"publicationState": "candidate"}Gate report
validate_candidate.py --document --text
Blocker: firm_attestation_missing. No authenticated firm signature; publisher claims are not firm attestations.
Result: stays a candidate. Every claim is AI-extracted and no firm has signed it, so it cannot be shown as firm-attested coverage.
02A real pass
One collection pass over public audit reports for the protocols in the explorer. Each binding names a file the report covers, with the page and a hash of the exact quote. None of it is released.
“The main logic of the protocol is in the Safe.sol contract.”
| Report | Protocol | Bindings | Checks passed | Candidate |
|---|---|---|---|---|
| Ackee BlockchainSafe smart account 1.4.0 audit | safe | 4 | 1 of 7 | JSON |
| CertoraEigenLayer redistribution slash delay v1.13.0 | eigenlayer | 2 | 2 of 7 | JSON |
| CertoraLido V3 audit | lido | 2 | 3 of 7 | JSON |
| ChainSecurityCode Assessment of the Euler Vault Kit Smart Contracts | euler | 34 | 3 of 7 | JSON |
| OpenZeppelinComet Asset Extension Audit | compound | 3 | 1 of 7 | JSON |
| ABDK ConsultingAave v3.5 code audit | aave | 3 | 2 of 7 | JSON |
| MixBytesAave v3.7 security audit | aave | 1 | 3 of 7 | JSON |
| BlackthornSecurity Review For Aave v3.4 | aave | 3 | 2 of 7 | JSON |
| CertoraAave v3.3 formal verification and manual review | aave | 1 | 1 of 7 | JSON |
| ABDK ConsultingAave V3 smart contract audit | aave | 1 | 1 of 7 | JSON |
| Trail of BitsAave V3 Security Assessment | aave | 1 | 2 of 7 | JSON |
| OpenZeppelinCompound III Audit | compound | 2 | 3 of 7 | JSON |
| CertoraEigenLayer Slashing (V1.0.0) formal verification | eigenlayer | 2 | 2 of 7 | JSON |
| CertoraSafe Smart Account v1.3.0 security review | safe | 1 | 1 of 7 | JSON |
| OpenZeppelinUniswap V4 Periphery and Universal Router Audit | uniswap | 1 | 2 of 7 | JSON |
Each cell is one check, in order: passed (filled), blocked (orange), not applicable (empty). Every report is blocked at the firm signature and the code match.
Result: in audit scope for the files named, cited and hashed. Code match not yet checked, and no release.
03
Each stage records what it used, so a reader can trace a wallet result back to the report text and the code it describes.
Published audit reports, the protocol's repositories and its deployment records. The auditor's own publication is preferred over a copy. A search result is a lead, not evidence.
origin.urlorigin.publisherThe report file is stored once, by its hash, with where it came from and when it was fetched. Later steps cite this exact file.
contentSha256fetchedAtDates, the reviewed commit, contracts and functions in scope, exclusions and findings are read out of the report. Every claim carries the exact quote and page it came from; a claim without a citation is dropped.
quotepagereviewedCommitEvery candidate goes through the same checks. Is the source the publisher's own? Is every quote in the document? Is the reviewed commit exact? Does the build reproduce, and does the deployed code match it? A step that is not done is recorded as a blocker, never skipped.
blockerscodeLinkage.statusImported evidence is released only after independent human review. Audits that an approved audit firm pushes from its own tooling are released by the firm's attestation plus automated code checks, with no AI on that path.
releaseIdreleasedAtFor each transaction, deterministic code compares the deployed code on the call path with released reviews. A part is Reviewed only when its exact code and function match a released review. Changed code reads Changed since review; anything without a released review stays Coverage unknown.
REVIEW_COVEREDCHANGED_SINCE_REVIEWREVIEW_COVERAGE_UNKNOWN04
Collection can speed up finding sources. It has no say in the result.
| Step | Decided by |
|---|---|
| Finding reports and sources | Decided by: AI-assisted collection and people |
| Turning report text into claims | Decided by: AI-assisted extraction, every claim cited and checked |
| Whether evidence is released | Decided by: Independent human review, or an approved firm's own attestation |
| Whether a transaction is covered | Decided by: The deterministic matcher, from released evidence only |
Until it is released, collected evidence appears as “Candidate evidence · not approved” and never counts as coverage in the Explorer or the wallet.
Approved audit firms push their audits by REST API or MCP from their own tooling. The release states whether the live code matches the audited commit.
AI-assisted collection is in development. Today evidence reaches the Explorer through human-reviewed imports and approved audit firms.