The verification gate

A claim is confirmed only if the test asserting the spec fails.

A claim is CONFIRMED only if the test asserting the documented contract FAILS.

If the test passes, the spec is being honoured and the claim was wrong.

How a claim is judged

1. Synthesise a request

   Built from the spec’s schema. Field names matter: a BVN that is not 11 digits gets rejected by validation before the response shape is reached.

2. Generate a Go test asserting the documented behaviour

   The test encodes what the spec promises, not what the code currently does, then runs against the real handler through httptest.

3. Run it

   Failure means code and contract disagree.

Verdicts

Verdict	Meaning	Reaches the report
confirmed	the test failed, real drift	Yes
refuted	the test passed, claim was wrong	No
unsupported	claim kind is not executable	No
error	the test did not build or run	No

A claim the gate cannot execute is marked unverifiable. It is never let through as though it had passed.

Executable claim kinds

response_field_mismatch · response_type_mismatch · status_code_mismatch · response_header_mismatch · auth_mismatch

Run it on one case

python3 auditor/verify.py eval/cases/D01

No model is involved. A claim is judged the same way whether a rule, an agent or a person raised it, so the same gate can score the baseline and the agent without favouring either.