Verification Receipts for AI Agents

Why verification needs its own object

Agent workflows need more than a binary payment event. A buyer may pay for access, escrow funds or issue credit before the work is complete. The system still needs a later record that says whether the delivered output matched the agreement.

The Verification Receipt is that record. It gives the verifier a structured way to say accepted, rejected or partially accepted, and to bind that verdict to the original agreement.

What a good receipt includes

A useful verification receipt should identify the agreement, the verifier, the verdict, relevant output references, timestamps and signature material. It should be deterministic enough for conformance testing and explicit enough for policy engines to evaluate.

The receipt does not magically make the verifier honest. It makes the verifier's claim inspectable.

Verifier design is a product decision

A verifier can be a deterministic test suite, a human reviewer, a model-based evaluator, a committee or an application-specific oracle. Accord does not prescribe one universal verifier. It gives each implementation a common receipt shape so downstream tools can reason about the verdict.

For production systems, verifier assumptions are part of the trust boundary and should be documented as carefully as payment or signing assumptions.

FAQ

Does a Verification Receipt prove the work is objectively correct?

No. It records a verifier's signed verdict under the agreement's verification rules.

Can work be partially accepted?

Yes. Accord's model allows a verifier to record accepted, rejected or partially accepted outcomes.