Concept
Approval Evidence
For each consequential action: exactly what the agent asked to do, which policy version decided it and which rule matched, who approved it and how they were verified, and whether it executed. The record has to be written when the decision is made, be tamper-evident, and be checkable without trusting the vendor that produced it.
What records do auditors expect for actions taken by AI agents?
For each consequential action: exactly what the agent asked to do, which policy version decided it and which rule matched, who approved it and how they were verified, and whether it executed. The record has to be written when the decision is made, be tamper-evident, and be checkable without trusting the vendor that produced it.
- Which controls does this evidence support?
- The ones you already report on: SOC 2 CC6 for who may act and approve, CC7 for monitoring what happened, and CC8 for changes to the policy itself. For AI systems in scope of the EU AI Act, it serves Article 12 on record-keeping and Article 14 on human oversight.
- Why aren't logs and screenshots enough?
- Auditors increasingly ask for proof that a process was enforced by the system rather than followed by a person. Evidence assembled afterwards — a screenshot of a settings screen, a message in a channel treated as sign-off, logs stitched together at audit time — can be edited, and it records what happened rather than what was permitted. A record produced by the check itself, while it ran, answers the question as it was asked.
- How does my auditor check the record without asking the vendor?
- Each Decision Dossier is signed with Ed25519. Your auditor verifies the signature offline against the published key set with the open-source @decionis/verify tool, so the check does not depend on our servers.
2 patterns answer this
Safe code deployment
A pipeline — or an agent with repo access — ships to production, force-pushes, or runs a migration.
1 policy pack
Policy change review
Someone edits a policy, raises a threshold, or disables a rule.
Its gate and signals
From this term to a governed action
- Safe code deploymentThe recurring action behind this term.
- DevOps & CI policy packThe pack whose rules decide it, rule by rule.
- Try to break it in the sandboxA real signed verdict, no account.
- Deploy on your surfaceAvailable on 3 platforms today.
- Verify the proof yourselfCheck the signature without trusting us.
- Start NowMeasure on your traffic before enforcing.
Where this fits
Decionis is the deterministic decision layer between intent and execution. This concept is one way in; the architecture behind it is the same in every case.