Every agent gets an identity, a risk tier and a policy before it is tested. ProofAgent then plays the adversary over many turns, a multi-jury evaluation with three independent jurors score what the agent actually did, and the result is sealed in a record a reviewer can verify.
How it works
Register the agent with its AI bill of materials and an EU AI Act aligned risk tier. Lock a governance profile. Evaluate under adversarial pressure, with code or through the agent's API; serious failures are re-tested. Gate the release as PASS, REVIEW or BLOCK, a CI exit code. Seal a Portable Evaluation Record with every finding, its quoted evidence and the rule it maps to.
What a record proves
The provenance and outcome of a recorded evaluation against your policy. It does not prove an agent is safe in every situation. Framework mappings organise evidence against obligations; they are not a certification.