Ask an AI coworker what it did last week and it will tell you something. The useful version of that answer is one you can check.
Every agent action here writes an evidence record: which agent, which command, what it read, what it produced, when, and against which configuration. Those records are then hash-chained — each record’s hash covers the previous hash plus a canonical serialization of the record itself. Change one field in one record from March and every hash after it changes.
Two honest caveats, stated here so the proof page doesn’t have to argue for them:
- The chain proves internal consistency, not external truth. It shows that the published segment is exactly what was anchored, and that nothing was quietly edited afterwards. It does not prove the anchor itself was honest — that requires the baseline to be independently visible, and the repository holding it is currently private.
- Signing is built but unprovisioned. The code that signs each record with an Ed25519 key is complete and tested. No keys have been generated, so the count of signed records is zero. The proof page reports that zero rather than omitting the column.
The chain is real. The signatures are not yet. Saying so is cheaper than being caught not saying so.