An AI audit log can show what a system recorded. But how can an independent reviewer verify that those records haven't been changed since they were created?
Traditional exports and screenshots don't provide that assurance on their own. Cryptographic hash chains, independently issued timestamps, and offline verification make it possible to check the integrity of an audit trail without relying on the system that produced it.
01 · The problem
What is in the export
When an auditor requests AI activity logs, they typically receive an export from the system that recorded them. Suppose the export contains 10,000 interactions. How can the auditor know that someone hasn't deleted 500 problematic interactions or modified records before sending the file?
A spreadsheet shows the information it contains, but provides no independent way to verify that the records haven't been altered.
Recordkeeping rules differ by regime: SEC Rule 17a-4 for broker-dealers, FINRA 24-09 for member firms, SR 26-2 for banks. An examiner asks the same thing: how do you know nobody edited these records?
02 · The mechanism
How independent verification works
Four pieces. Once the vendor hands over the files, the reviewer checks everything alone.
- A tamper-evident chain. The system seals each record with a hash, a short fingerprint of its contents, and each hash also covers the hash of the record before it. Edit one character, or delete a record from the middle, and the chain breaks there.
- An export of the records. The vendor sends the review period's records as files the reviewer keeps and can check again later.
- An independent verifier. The reviewer runs a program on their own computer that recalculates every hash and checks the links between records. It works offline, without accessing the vendor's systems. The source code is public, so anyone can inspect how the verification works.
- Independent timestamps. Periodically the system sends the chain's fingerprint to an outside timestamp authority, which signs it with the current time. A chain alone only proves the records agree with each other: whoever holds them could rewrite everything and seal a new chain. A rewrite produces a different fingerprint, which will not match what the authority signed.
What this proves. A passing check answers one question: has anyone changed these records since they were sealed? It does not tell you the records were accurate when written, or that the export contains every record. The chain cannot reveal an AI call that nobody recorded.
Where completeness comes from. Meilynx records each interaction as part of enforcing its policies, so governed interactions are captured automatically. No audit trail can account for AI activity that bypasses the governance system.
03 · The flow
From record to result
The check runs on the reviewer's side of the handoff.
01 · Written
An AI interaction is written down as a record.
02 · Sealed
The record gets a fingerprint that also covers the record before it. A later edit breaks the link.
03 · Anchored
The chain's fingerprint goes to an outside authority, which signs it with the time and returns a token.
04 · Exported
The records for the period go to the reviewer as files, with the tokens that cover them.
05 · Verified offline
A verifier whose code the reviewer can read recomputes every fingerprint and link, and checks each token.
06 · Result
Intact. Every record matches its seal.
Broken. The check names the record where the chain fails.
04 · In practice
Putting it into practice
As the compliance officer, internal auditor or third-party risk reviewer, ask the vendor for four things:
- the records for the review period, as an export you keep;
- the timestamp tokens covering those records;
- a verifier you can run yourself, with published source code;
- a written statement of which AI traffic the system records, so a passing check is not mistaken for full coverage.
Banks can include these evidence requirements in vendor contracts through audit and reporting provisions, and incorporate them into ongoing third-party risk monitoring. AI vendors serving financial institutions should be prepared to provide this evidence as part of the due diligence process.
Try it on real evidence. Request our sample examination evidence: a simulated audit chain plus the open-source verifier, which runs offline on your machine. Edit a record, run the check again, and watch it fail.
Frequently asked
Does verifying an audit chain prove the log is complete?
No. The chain covers the records inside it and shows that nobody changed them after sealing. It cannot reveal an AI call that nobody recorded. Completeness depends on where records get written: a system that writes each interaction while it enforces its policies captures those interactions automatically, and AI activity that goes around that system leaves nothing to verify.
Does SEC Rule 17a-4 require WORM storage for AI records?
Since 3 May 2023, a broker-dealer can meet Rule 17a-4(f) in either of two ways: keep records in a format nobody can rewrite or erase, or keep a complete time-stamped audit trail of every change and deletion that allows the original record to be re-created. An AI output that counts as a required record goes into a system that does one of the two.
Why should an audit trail verifier run offline?
An offline check depends only on the exported files and the verifier's code. No vendor system takes part, so the reviewer can repeat the check later on the same files and get the same answer.