Skip to content

Accountability infrastructure for autonomous engineering

Principle 03

Evidence over claims

Every consequential outcome is bound to an artefact that can be re-examined independently: a commit, a prepared-state digest, a durable transition or a signed receipt.

Prose summaries decay, get rewritten and cannot be re-checked. Evidence is bound to an exact commit or prepared state and can be inspected long after the session that produced it has ended.

Bind evidence to exact state

A summary that says the tests passed is unfalsifiable after the fact. A record that says which commit was built, what the run produced and what the verifier observed can be checked again tomorrow, by someone who was not there.

AgentProof binds authority to an exact prepared state digest. If the state drifts between approval and execution, the authority no longer applies — because the thing that was approved no longer exists.

Durability outlives the session

Agent sessions are ephemeral. Context windows are discarded, terminals close, logs rotate. Any accountability model that lives inside the session disappears with it.

DoneState writes transitions durably before the effects they describe, so recorded state is never ahead of reality and a crashed run resumes to a known point instead of a guessed one.

Verifiable without privileged access

A receipt that can only be checked by the system that issued it is a log entry. AgentProof receipts bind action, prepared state, authority and signer so that verification needs the receipt and a public key — not credentials for the system under scrutiny.

This is the difference between an audit trail you are asked to believe and an audit trail you can check.

Give agents authority. Keep the proof.

Principles are only credible when the architecture makes them expensive to violate.