Skip to content
· 5 min read

Why agent payments need an append-only audit trail

An AI agent audit trail has one job that a log cannot do: let a party who does not trust you check it. That single requirement changes the design, because a log is written for the operator and an audit trail is written for a stranger.

Tamper evident, not tamper proof


Google and Mastercard have been co developing a standard called Verifiable Intent, described as a tamper proof log of user authorised agent actions, and both donated their work to the FIDO Alliance in 2026. The phrase is doing a lot of work. Nothing in software is tamper proof. A database administrator can drop a table. An operator can edit a row. A migration can rewrite a partition. The honest engineering claim is narrower and stronger: tamper evident, meaning the evidence of an edit is detectable by anyone holding the chain.


That is the standard I hold myself to. Each record carries the hash of the record before it. Change one row in the middle and every hash after it no longer matches. You do not need to prevent the edit to get value from it, you need the edit to be impossible to hide from someone who was not part of it.

Why append only, specifically


Because the alternative is a history that can be revised, and a revised history cannot answer a question about what was true at the time. Append only is not a moral preference, it is the only structure that lets a record keep meaning after the facts around it have moved on. A correction is a new row that references the row it corrects. The old row stays, with the same hash it had, and the reader sees both.


The payment networks already require something in this direction. Visa's Acceptance Risk Standards oblige acquirers to retain dispute investigation detail for a minimum of two years, including the event description and analysis. Two years of retained investigation means two years during which somebody may ask what the system knew. An append only structure is what makes that answer possible.

What the audit trail is actually for


  • Disputes. The clearest writing on this comes from a payments law firm: the party with the better record does not merely win, it controls the framing and whether the fight reaches a court.
  • Reconciliation. Every decision has to tie to a real charge or a real refusal, and the mismatch has to be findable by comparing two systems rather than one.
  • Errors. A bad instruction produces a pattern across rows long before it produces a complaint. Append only is what lets you query the past at all.
  • Obligation. If a regulator asks who authorised a payment, the answer should not require your cooperation to produce.

The fourth is the one that changes how you build. If the regulator is a plausible future reader, then the record is not an internal artefact and the chain has to survive export. A trail that only exists in your database is a log with extra steps.


The verification is a procedure, and it should be written down as one, because a property nobody can check is a slogan. Take any row. Recompute its hash from the fields it claims to contain, including the hash of the previous row. Compare it with the stored hash. Then walk the chain backwards from the newest row and confirm every link. Three steps, no access required beyond the export, no permission from us. That is the test I would want a third party to run against me, which is why I publish the chain in a form that can be exported without asking.


The same procedure runs in the suite on every commit, so a change that breaks the chain fails the build rather than a customer's dispute. That is the only sense in which I would use the phrase verified by test, and it is deliberately a small sense. It does not mean the ledger cannot be wrong. It means a specific, checkable property holds, and I can tell you exactly which one.

The part that is hard: deletion


European data protection law gives people a right to erasure, and append only looks like its opposite. The resolution is to separate the payload from the chain. The row keeps its position, its timestamp, its hash and its links. The personal payload is removed. A reader sees that a row existed, that it was redacted, when, and under what instruction. The chain stays intact and the history stays honest.


I would rather have that argument with a regulator than have a silent gap in a chain. A gap is indistinguishable from tampering to exactly the person who is looking for tampering.

What I got wrong, and what the audit found


I commissioned an external audit of the money path earlier this year, and one of the findings was a ledger that could lie to itself: a read path that trusted a cached total instead of recomputing from the entries, which meant a bug in the cache would produce a number that looked authoritative and was not. The fix was unglamorous. The read path recomputes, or it says it does not know. An audit trail that can be wrong in a way nobody can detect is not an audit trail, it is a source of confidence.


So the claims I make about this are narrow on purpose. The chain is verifiable, and the verification is a procedure anyone can repeat. That is not the same as saying the record is perfect, and I would rather make the smaller claim.


I am building it at Noesia. The fields that go into each row are in what an agent authorization record should contain, the property that makes the chain recomputable is in deterministic authorization for AI agents, and the reason the dispute is the forcing function is in who is responsible when an AI agent buys the wrong thing. What an external audit found in my ledger is what a security audit taught me.

Building Know Computer, a personal context layer for the AI era. Reply via [email protected].

More

Notes