Auditor pack
This page is for auditors who test system-generated evidence: SOC 2 and ISO examiners, internal audit, regulators, and litigation-support teams. It covers how to get the full population, how to test it for completeness and accuracy, which obligations each artifact is designed to support, and a template a customer's qualified person can adapt. For AI reviewers, the machine-readable version is auditor.json.
1. Population, completeness and accuracy
| Test | How | Where |
|---|---|---|
| Full population | Every proof each validator cluster has written, with its time, verdict and hash. Export it as CSV, or read the signed index directly. | Incidents → Export the full population · /evidence/hourly-proof/index.json · /evidence/hourly-proof-mr02/index.json |
| Completeness: nothing removed | Each proof carries the SHA-256 of the one before it, and the transparency log's Merkle tree must contain every earlier checkpoint. Removing, editing or reordering a proof breaks both. | Transparency log · verify_hourly_proofs.py |
| Completeness: nothing missed | Proofs are numbered without a gap, and every stretch longer than 45 minutes with no proof written is listed with its length. | Incidents → Completeness |
| Accuracy: the record is what was signed | Recompute each record's SHA-256 and check its Ed25519 signature against the pinned key. Signatures on the quorum certificates are checked against the member keys pinned in the signed index. | Auditor console · Verification run |
| Continuous monitoring, by you | Run the independent log monitor hourly on your own machine or in your own CI. It keeps its own copy of every checkpoint and raises ALARM on a fork or rewrite. | verify_tlog_monitor.py · GitHub Actions workflow |
| Point in time | Each transparency-log checkpoint is timestamped by two outside RFC 3161 timestamp authorities (FreeTSA, DigiCert), and the cain-mr-01 chain head is anchored daily in Sigstore Rekor. | Transparency log · Rekor anchors |
| Customer decision records | The customer exports a record with /fabric/decisions/<id>/signed-record and runs verify_decision_record.py. It checks ENCODING, DIGEST (recomputed), KEY and SIGNATURE. A changed verdict fails DIGEST, and a changed verdict with a recomputed digest fails SIGNATURE. | Verify your own evidence |
| Change history of the evidence pipeline | The repository commit and deployment that produced the current state, the dated changelog, and the known-limitations list. | /now.json · changelog |
2. Control-to-evidence map (self-assessed)
Our own mapping of artifacts on this site to record-keeping obligations and frameworks, so an auditor can find the evidence for each one quickly.
A conformity assessment, certification or legal opinion. No harmonised standard for EU AI Act logging is cited yet, so nobody can claim presumed conformity with Art. 12. No accredited body or CPA firm has assessed CAIN-42.
| Obligation or framework | What it asks for (summary) | Evidence here, designed to support it |
|---|---|---|
| EU AI Act Art. 12: record-keeping | Automatic logging that allows traceability (risk situations, post-market monitoring, deployer monitoring) | Signed, hash-chained decision records per agent action; per-tenant listing and export; transparency log |
| EU AI Act Art. 19 and Art. 26(6): log retention | Providers and deployers keep logs for at least six months | Append-only records with no deletion path; retention is set per deployment and is not yet a configurable control on this site |
| EU AI Act Art. 72: post-market monitoring | Systematic collection and analysis of performance data, including interaction with other AI systems | Continuous verification every 15 minutes (failures included), incident history from signed proofs, agent-run (trajectory) views |
| EU AI Act Art. 73(6): serious incidents | Do not alter the system in a way that affects later evaluation before informing authorities | Records cannot be edited without detection (hash chain, signatures, checkpoints, outside timestamps); a dedicated "frozen snapshot" control is not built yet |
| NIST AI RMF 1.0: MEASURE and MANAGE | Track, measure and manage AI risks with documented evidence | Signed claims register with statuses and limits, tamper lab, live decision trials, known-limitations list |
| US FRE 902(13)/(14) | Self-authentication of machine-generated records and copies via certification by a qualified person (hash comparison) | Published SHA-256 for every file, signed records, the template below |
3. Certification template for your qualified person (FRE 902(13)/(14))
A starting point that a customer's own qualified person (for example an IT custodian or forensic examiner) can adapt, sign and give to counsel.
This is not legal advice, and FIRME AI CORPORATION does not certify anything here. Whether a record is admitted, and what weight it gets, is decided by the court. Hearsay and other objections are separate questions. Have counsel review the final certification.
CERTIFICATION OF RECORDS GENERATED BY AN ELECTRONIC PROCESS OR SYSTEM (Fed. R. Evid. 902(13) / 902(14)), DRAFT FOR COUNSEL REVIEW I, [full name], certify as follows: 1. I am [title] at [organization]. I am familiar with how [organization] uses the CAIN-42 runtime and how its decision records are produced and kept, and I am qualified to make this certification. 2. The attached records ([list: decision ids / bundle file names]) were produced by the CAIN-42 runtime for tenant [tenant id] in the regular course of [organization]'s activity. 3. Collection. On [date, time, timezone] I downloaded each record from [URL, e.g. https://cainstudio.online/fabric/decisions/<id>/signed-record] using our API key. I computed its SHA-256 at collection: [hash per record, or attached hash list]. 4. Verification. On [date] I ran the published verifier verify_decision_record.py (SHA-256 of the verifier file: [hash]) against each record with the decision-signing key fetched from [URL] (key id [id]). It reported ENCODING, DIGEST, KEY and SIGNATURE as PASS for every record. [For evidence bundles: verify_keystone.py with the issuer key pinned by fingerprint [SHA256:...].] 5. Before this certification was signed, I computed the SHA-256 of each copy again and it matched the collection-time hash in paragraph 3. The copies are therefore identical to the records as collected. 6. [Optional] The records fall inside transparency-log checkpoint [origin, size, root] of clawx.click/tlog, timestamped by [TSA] at [time] (token attached). I certify under penalty of perjury that the foregoing is true and correct. Signature: ______________________ Date: ____________ Place: ____________
Copy the text above, or use auditor-902-template.txt.
4. What an auditor cannot get here yet
- An independent third-party audit report. None has been completed.
- Independent witness cosignatures on the transparency log. There are 0 so far, and the log has not yet been accepted by a witness network.
- A SOC 2 report, an ISO/IEC 42001 certificate, or eIDAS qualified timestamps. The timestamps here are RFC 3161 tokens from FreeTSA and DigiCert.
- Validators run by anyone other than FIRME AI CORPORATION.
Something here doesn't hold up? Report it. Corrections are published as new, superseding evidence, never as silent edits.