“Six Months of Proof: Independently-Verifiable Records for Agent Actions Under the EU AI Act”

In five lines

1. The EU AI Act asks for automatic records of what high-risk systems do, kept for at least six months: only if the system is classified high-risk, from 2 Dec 2027 for Annex III systems and 2 Aug 2028 for Annex I.
2. A law can require that logs exist. It cannot make the party that keeps them an impartial witness.
3. Anchoring fixes that: seal each record to a digest and write the digest to an independent append-only log.
4. The Agent Action Capsule project does this on open IETF transparency-log standards (SCITT, COSE), so anyone can check a record without the operator's help.
5. Live demo: emit a record at an action boundary, anchor it, check it with the open verifier, then change one byte and watch the check fail.

The obligations behind the talk are mapped item by item on EU AI Act: what a record can show. The report from the talk, with two more worked examples, is on the demos page.