Abstract
A wallet user who is about to sign a transaction usually gets a protocol name and, at best, a link to an audit PDF. The PDF covers one commit and one scope. The contract that will run may be a proxy pointing at a different implementation, may have been upgraded since the audit, or may never have been in scope. Nothing at the moment of signing tells the user which of these is true.
Proof Before Sign answers one narrow question for each contract call inside an unsigned transaction: which public audit, if any, covers the code that will actually execute, and does the deployed bytecode equal the audited build. The answer is computed by a hosted analyzer, signed with an Ed25519 key held in Google Cloud KMS, and verified inside a MetaMask Snap against a published trust file before it is shown.
Today the system checks real mainnet transactions on ten protocols, has rebuilt two deployed contracts byte for byte from their audited commits, and labels every unknown as unknown. It has no released audit yet, because a release needs two independent human reviewers and none have been named. This grant funds the open release of the verifier and evidence format, the first released audits, public Snap distribution, and adoption by outside wallets and agent executors.
1. The problem
Audits are published per commit and per scope. Users interact with addresses. Between the two sit several steps that nobody shows at signing time:
- Proxies. The address in the transaction is often a proxy (ERC-1967, beacon, Diamond facets). The code that runs is the implementation it delegates to, which can be replaced.
- Upgrades. An implementation audited in March can be replaced in June. The audit link on the protocol's site does not change.
- Scope. A report names specific files. Libraries, periphery contracts and later extensions are often outside it.
- Build provenance. Even when the repository and commit are right, the deployed bytes may come from a different compiler, settings or revision.
These are not hypothetical. Rebuilding the Uniswap V3 factory from the tag ABDK reviewed gives a different runtime hash from the deployed factory, and Morpho Blue's deployed bytecode rebuilds from a revision later than the one OpenZeppelin reviewed. Neither difference is a vulnerability. Each is a gap in provenance that a user cannot see, and that an attacker who replaces or wraps audited code can rely on.
AI agents make the gap larger. An agent that pays or signs on a user's behalf needs a machine-readable answer to "is the code I am about to call the code that was audited?", and today there is no such answer to consume.
2. The question we answer, and the ones we do not
For one unsigned transaction at one pinned block, Proof of Audits reports, for every contract call on the execution path:
- which contract and function actually executes, after resolving proxies;
- which public audit report, if any, names the source file of that code, with a quoted citation from the report;
- whether the deployed runtime bytecode equals a rebuild of the commit the report names;
- whether that evidence has been released by two independent human reviewers.
It does not say a contract is safe, does not rate auditors' opinions, does not hold funds or keys, and does not block a transaction. A wallet shows the evidence; an agent's own executor decides what to do with it under rules its owner sets.
3. The evidence ladder
Every call gets exactly one status. The statuses are ordered by strength of evidence, and the interface never colours anything below a release green.
| Status | What it means | What it does not mean |
|---|---|---|
| Not checked | The code is outside the ten catalogued protocols, or its source file could not be established. | That the code is unaudited. |
| No audit found | We hold the protocol's reports with extracted file lists, and none names this file. | That the code is unsafe. A report we have not collected may cover it. |
| In audit scope | A collected public report names this source file. The citation (page, quote, quote hash, document hash) is attached. | That the deployed bytes equal the reviewed code. The deployed code has not been compared. |
| Byte-for-byte match | The deployed runtime bytecode equals a rebuild of the exact commit the report names, at a pinned block. | That the evidence is released. Two independent reviewers must still approve it. |
| Released | Two distinct human reviewers approved the match and its scope, and an operator signed the release. | That the code is safe. A release records what an audit covered, not that the audit found everything. |
4. How it works
4.1 Pipeline
- Intercept. The MetaMask Snap receives the unsigned transaction through MetaMask's transaction-insight hook, before the confirmation screen is shown. Nothing is signed or broadcast by Proof of Audits.
- Pin and trace. The analyzer pins a block and simulates the call with
debug_traceCalland the call tracer on a tracing-capable Ethereum RPC. If no trace is available, the result is UNKNOWN with the decoded function kept, never a guess. - Resolve code. Each call frame is resolved to the code that ran: proxies are followed (ERC-1967 and beacon slots are read at the pinned block), and the runtime code hash of every executing address is recorded.
- Match evidence. Each frame is matched against three evidence sources in order: released records, exact-build records (runtime hash equals the hash of the audited rebuild) and cited scope bindings (a report names the source file, established through the verified source paths of the deployed code).
- Sign. The analyzer signs a runtime proof (payload type
application/vnd.poa.runtime-proof.v1+json) with keypoa-google-runtime-v1, an Ed25519 key in Google Cloud KMS. - Verify and show. The Snap verifies the signature against the trust file shipped in its bundle, checks freshness, and renders the panel inside MetaMask's own confirmation. An unsigned or wrongly signed response is shown as a verification failure, not as evidence.
4.2 Trust file and keys
| Field | Value |
|---|---|
| Trust domain | poa-production-ethereum-v1 |
| Policy versions | poa_policy_v2, poa_policy_v3 |
| Signature threshold | 1 |
| Key id | poa-google-runtime-v1, version 1, Ed25519 |
| Public key | a16875405345d91554b9539d265c051b5365b7ed6b3ed99c4209428213a7a062 |
| Valid | 6 October 2026 to 7 October 2027, chain 1 (Ethereum mainnet) only |
| Private key | Never leaves Google Cloud KMS (keyRings/poa-proof-v1/cryptoKeys/runtime-ed25519) |
A transaction passport expires ten minutes after analysis and is held only in memory. It is evidence about one transaction at one block, not a reusable safety label for a contract. Key rotation works by publishing a new key in the trust file with an activation time and retiring the old one, and the Snap rejects proofs signed outside a key's window.
4.3 Exact build matching
A byte-for-byte match is established without trusting the explorer's "verified" badge. The reviewed commit is taken from the audit report itself and resolved to a full Git hash. The commit is compiled with the pinned compiler binary and the project's committed settings. The resulting runtime is compared with eth_getCode at a pinned block, from more than one RPC provider. Only documented transformations are allowed: immutable values written at deployment, linked library addresses, and compiler metadata when the project's settings strip it. Anything else is a mismatch.
4.4 Verdict vocabulary
| Verdict | Meaning |
|---|---|
| VERIFIED | Released audit evidence covers the exact live code and the path this transaction needs. |
| PARTIAL | Evidence covers part of the interaction; something material is not covered. |
| CHANGED | Live code, path, configuration or control differs from what was reviewed. |
| NOT VERIFIED | No released evidence applies, or the facts could not be established. Not a finding of risk, not a sign of safety. |
Per call, the component states are REVIEW_COVERED, CHANGED_SINCE_REVIEW, EXCLUDED_FROM_REVIEW, REVIEW_COVERAGE_UNKNOWN, UNRESOLVED, NO_CODE (a wallet or precompile) and PROXY_ENTRY (the forwarding proxy, whose implementation is what gets assessed). Transaction coverage is FULLY, PARTIALLY or NOT_REVIEW_COVERED, INCOMPLETE when the path could not be resolved, or UNKNOWN. Unknown and incomplete are grey, never green.
4.5 Release: two humans, enforced in the database
AI is used only to collect candidate evidence (find reports, extract file lists, propose quotes). Every quote is checked against the document text and its hash, and a validator records the result of eight gates for each candidate; a failed or blocked gate is never upgraded to coverage (section 4 of the audit-gap map whitepaper). An AI step cannot approve anything.
A release goes import, verify, review, release. Review requires the EVIDENCE_REVIEWER role, and the database itself refuses to release unless there are exactly two approvals from two distinct reviewers, neither of them the importer (migration 006_evidence_release.sql raises matching independent approvals required). An operator then signs the release. Audit firms that push their own reports through the firm API are attributed to the firm and still get the automated bytecode check, whose result is shown next to the release.
4.6 AI agents
The same signed evidence is available read-only through the API and an MCP server. An agent's executor, controlled by the agent's owner, decides whether to allow, block or ask for approval. A local reference executor passes 15 checks on Anvil (spending caps, destination, selector, code-hash binding, replay and revocation) using owner-supplied fixtures. It does not yet verify Proof of Audits signatures and runs in no outside production system; milestone 4 changes that.
5. Evidence today
5.1 Real transactions, signed by the production analyzer
| Transaction | Block | Calls | Byte-for-byte match | In audit scope | No audit found | Not checked |
|---|---|---|---|---|---|---|
| EigenLayer · restake 1 stETH (StrategyManager) | 26,404,864 | 23 | 0 | 19 | 4 | 0 |
| Euler V2 · deposit 50 USDC into eUSDC-2 through the EVC | 26,404,835 | 15 | 0 | 13 | 2 | 0 |
| Aave V3 · supply 100 USDC | 26,404,833 | 14 | 0 | 12 | 2 | 0 |
| Aave V4 · supply 1 WETH | 26,403,765 | 3 | 1 | 0 | 2 | 0 |
| Sky · deposit 100 USDS into sUSDS | 26,404,834 | 9 | 0 | 9 | 0 | 0 |
| Uniswap V4 · swap 0.1 ETH for USDC (Universal Router) | 26,404,835 | 11 | 0 | 9 | 2 | 0 |
| Uniswap V2 · approve the USDC/WETH LP token | 26,403,765 | 1 | 1 | 0 | 0 | 0 |
| Lido · stake 0.1 ETH | 26,403,766 | 4 | 0 | 4 | 0 | 0 |
| Safe · execute a 0 ETH transaction to the Safe itself | 26,404,835 | 4 | 0 | 4 | 0 | 0 |
| Compound III · supply 100 USDC to the USDC Comet | 26,404,834 | 8 | 0 | 2 | 6 | 0 |
| Morpho Blue · authorize another address | 26,403,765 | 1 | 0 | 1 | 0 | 0 |
| Morpho Blue · supply 50 USDC to the wstETH/USDC market | 26,404,834 | 4 | 0 | 1 | 3 | 0 |
12 checks across 9 protocols, blocks 26,403,765 to 26,404,864: 97 contract calls, of which 2 byte-for-byte matches, 74 in a cited audit scope, 21 with no audit found and 0 not checked. Each response was signed by the production analyzer and verified against the published trust file.
Each row is browsable at https://proofofaudits.com/demos/proof-before-sign, call by call, with the report citation behind every "In audit scope" and the rebuild record behind every match. Aave V3 supply is a good example: 12 of its 14 calls sit in a cited audit scope (MixBytes, ABDK, Blackthorn, Certora, Trail of Bits). The other two are the USDC token proxy and implementation: Circle says Halborn audited USDC v2.2, but that report is not published, so they show No audit found.
5.2 Byte-for-byte matches
| Contract | Address | Audit and commit | Runtime keccak-256 | Released |
|---|---|---|---|---|
| Uniswap V2 Pair (USDC/WETH) | 0xb4e16d0168e52d35cacd2c6185b44281ec28c9dc | dapp.org audit, v2-core 8160750, solc 0.5.16, 999,999 optimizer runs, 11,293 bytes | 0x5b83bdbcc56b2e630f2807bbadd2b0c21619108066b92a58de081261089e9ce5 | No (0 of 2 reviewers) |
Aave V4 Spoke implementation (behind proxy 0x94e7a5dcbe816e498b89ab752661904e2f56c485) | 0xabd0e26fe17bde4f1f1187ed8aa80c274e03d8b5 | ChainSecurity, 23 March 2026, final reviewed commit cdacec5, solc 0.8.28, 24,291 bytes | 0x41b906d98443c68139467f7bfd8194e361c8d488c55a3cded7ecbf30c9b42eb6 | No (0 of 2 reviewers) |
Both runtime hashes were re-read from Ethereum mainnet on 13 November 2026 and still match. The Aave V4 proxy frame itself is outside the ChainSecurity scope and is reported as such.
5.3 Software
- MetaMask Snap
@proofofaudits/proof-before-sign-snap, published on npm (0.7.x). It runs in MetaMask Flask. It is not yet on MetaMask's Snaps allowlist, so stable MetaMask cannot install it. - Hosted analyzer on Google Cloud Run (
POST /api/proof/analyze), returning signed, transaction-bound responses. - Public explorer of 656 callable functions on ten protocols with the evidence for each: https://proofofaudits.com/explorer/functions.
- 960 automated tests in the
@poa/proof-before-signpackage (110 files, exit 0 at commitb98adfae). - A Chrome extension (1.0.0) in developer preview only, in no store; 65 tests passing at its preview commit.
6. Threat model and failure modes
| Failure or attack | Effect | Response |
|---|---|---|
| Tracing RPC unavailable or rate-limited | No execution path | Signed UNKNOWN, decoded function kept. Never a coverage claim without a trace. |
| RPC returns false code or state | Wrong match | Block anchor and code read from two independent providers; milestone 3 makes a second trace provider mandatory. |
| Contract upgraded after release | Old evidence no longer applies | The code hash is bound into every record. A new hash invalidates the match and the call shows CHANGED. |
| Analyzer response tampered in transit | False evidence shown | Ed25519 signature verified inside the Snap against the bundled trust file; failure is shown as failure. |
| Signing key compromised | Forged proofs | Key lives in KMS; trust file supports rotation with activation and retirement times; the Snap rejects keys outside their window. |
| A protocol pays for a better verdict | Corrupted evidence | Commitment: payment never changes what the evidence says, and no one can buy a coverage verdict. |
| AI extraction invents a scope | False In audit scope | Every claim needs a quote whose hash matches the retrieved document; AI cannot approve; release needs two humans. |
| Reviewer collusion | False release | Two distinct reviewers who are not the importer, both recorded with the release; releases and their evidence are public and reproducible. |
7. Known limits, stated up front
- No released audit exists today. The two matches await two independent reviewers.
- Stable MetaMask users cannot install the Snap until MetaMask allowlists it.
- "In audit scope" is file-level evidence from a report. It is not proof that the function was reviewed.
- The explorer's dependency view is static source analysis at a pinned block, not a live trace.
- Ethereum mainnet only.
- No outside adopter, funder or revenue exists today.
8. What the grant funds
An open evidence format and an open verifier any wallet or executor can call, with the Snap as the first client: audited commit, build, deployed bytecode, proxy route and changes since the audit, shown before signing. Proposed ask: $75,000 over 12 months, of which $25,000 (one third) is tied to adoption.
| # | Milestone | Budget | Pass condition for the independent reviewer |
|---|---|---|---|
| 1 | Open-source release of the verifier, evidence schema, Snap and explorer data. Code under Apache-2.0, data under CC BY 4.0. | $15,000 | Public repository; a third party reproduces the Aave V4 and Uniswap V2 matches from the documentation alone. |
| 2 | First released audits: at least 3 mainnet contracts released through the two-reviewer process, each with build match and proxy route. | $20,000 | Releases visible in the public explorer and verifiable from the published evidence. |
| 3 | Snap on stable MetaMask (Snaps Directory) and a mandatory second trace provider. | $15,000 | Snap installable in stable MetaMask; verdicts reproduced on two independent RPC providers. |
| 4 | Adoption: at least 1 outside wallet or explorer shows the verdict in production, and at least 250 weekly active checks. | $25,000 | Partner integration live; weekly check counts published from public logs. |
Adoption outreach targets (none committed yet): MetaMask, Rabby, Trust Wallet and Ledger for wallets; Etherscan and Blockscout for explorers; agent-wallet teams for the executor gate.
Commitments
- Basic wallet evidence stays free for individuals.
- Auditors are never charged to submit evidence.
- Payment never changes what the evidence says.
- Synthetic and local results stay labelled as such.
- Released records and their reproduction scripts are public.
9. Check it yourself
The deployed half of both matches can be checked with nothing but a public RPC and Foundry's cast:
RPC=https://ethereum.drpc.org
# Uniswap V2 USDC/WETH Pair: expect 0x5b83bdbc…9e9ce5
cast keccak $(cast code 0xb4e16d0168e52d35cacd2c6185b44281ec28c9dc --rpc-url $RPC)
# Aave V4 Spoke implementation: expect 0x41b906d9…b42eb6
cast keccak $(cast code 0xabd0e26fe17bde4f1f1187ed8aa80c274e03d8b5 --rpc-url $RPC)
# Proxy route: the ERC-1967 implementation slot of the Spoke proxy
cast storage 0x94e7a5dcbe816e498b89ab752661904e2f56c485 \
0x360894a13ba1a3210667c828492db98dca3e2076cc3735a920a3ca505d382bbc --rpc-url $RPCThe audited half: clone Uniswap/v2-core at 816075049f811f1b061bca81d5d040b96f4c07eb, compile UniswapV2Pair with solc 0.5.16, optimizer on, 999,999 runs, EVM version istanbul (the repository's Waffle settings), and hash the runtime bytecode. The dapp.org auditors' own formal model embeds the same bytes (lemmas.k.md). The full scripts, receipts and hashes for both matches are published with milestone 1.
The signature on any recorded check can be verified with the public key in section 4.2. The recorded checks are at https://proofofaudits.com/evidence/demos/analyses/.
10. Team, reviewer and co-funding
- Team. Proof of Audits, led by its founder, Veer (veer@proofofaudits.com). Track record is the working system above: the signed analyzer, the Snap on npm, the ten-protocol explorer and the two byte-for-byte rebuilds, all checkable from the links in this paper.
- Independent technical reviewer. To be agreed with TheDAO Security Fund before milestone 1 starts. We ask for a reviewer with no ties to Proof of Audits, the ten catalogued protocols or the audit firms cited.
- Co-funding. None committed as of 13 November 2026. We will ask the wallets and protocols that benefit (the ten catalogued protocols and the wallet targets above) to pledge about 15%.
References
- ChainSecurity, Aave V4 code assessment, 23 March 2026https://github.com/aave/aave-v4/blob/main/audits/2026-03-23_Aave-V4_ChainSecurity.pdf
- dapp.org, Uniswap V2 audit reporthttps://dapp.org.uk/reports/uniswapv2.html
- Uniswap v2-core at the audited commit 8160750https://github.com/Uniswap/v2-core/tree/816075049f811f1b061bca81d5d040b96f4c07eb
- Aave V4 at the audited commit cdacec5https://github.com/aave/aave-v4/commit/cdacec509e4f848bff1a5556f503afa83eee3b79
- ERC-1967: Proxy storage slotshttps://eips.ethereum.org/EIPS/eip-1967
- MetaMask Snaps: transaction insightshttps://docs.metamask.io/snaps/features/transaction-insights/
- Proof Before Sign Snap (public source)https://github.com/ProofOfAudits/proof-before-sign-snap
- TheDAO Security Fundhttps://thedao.fund/