Verification checks
The evidence reference: each check the receiving client runs automatically, the references it uses and where it runs, for integrators and auditors.
Standard Proof is execution evidence that the receiving integration verifies against its own accepted policy before it uses a result. The receiving client runs these checks automatically before it delivers a result, so the next agent can use it without manual inspection. This page lists the checks for integration testing, auditing and troubleshooting.
Input bindings identify the artifacts or snapshots used. Source provenance is covered only when the operation includes evidence of the source. Recipients evaluate source content under their own acceptance policy.
Where verification runs
The recipient's integration, SDK, or verification software checks the evidence in an environment it trusts. A dashboard status or an agent's success message is not a substitute for those checks.
Configure the dependent action to require successful verification. Displaying a proof status to the model does not enforce that requirement.
The same requirements apply when the same agent continues its own work, not only when one agent hands work to another. An agent that requested an operation and then acts on its result still proceeds only through an integration that verifies Standard Proof, evaluates the result and enforces authority for the next action.
Accepted references must be established independently of the candidate result. They include the runtime identity and revocation state, evidence-signing trust roots, workload and input digests, applicable authority, and the recipient's verification policy.
Claims to check
The required evidence depends on the selected operation and execution boundary.
| Claim | Required relationship |
|---|---|
| Runtime | Attestation must match the accepted environment and applicable freshness or revocation requirements. |
| Workload | The recorded workload must match the approved code, model, or verifier. |
| Inputs | Recorded input bindings must match the expected artifacts or snapshots. |
| Authority | The operation and release must be covered by the required permissions. |
| Result | The evidence must bind to the exact returned result and intended release context. |
| Signature | The signed evidence must validate under the accepted trust model. |
Missing evidence for a required claim is a verification failure. A claim outside the recorded boundary remains unverified. Cleanup is reported separately and is not proof of physical erasure.
Conversation analytics outcomes
The Conversation analytics example distinguishes four outcomes: counts ready, an input check failed, publication held, and invalid evidence. Duplicate conversation IDs are checked by the example function. An accepted failure report contains no category counts; it is not a successful classification.