faffter-noon-evaluate

Run a holdout evaluation that exercises the spec's born-verifiable done-criteria against the running feature's endpoints. The evaluator classifies each criterion as scenario, assertion, or prose, judges only the first two against live responses, and emits a holdout-verdict contract block.

It is an evaluation where the judge receives only the spec and a running environment, never the diff, codebase, or build history. Blindness is attested by an external spawner that provably withholds the repository from the evaluator process, so the verdict cannot be influenced by the implementation.

Prose criteria are never machine-judged; they are mechanically forced to a needs-human verdict with no evidence flag. The holdout-verdict contract rejects any prose criterion judged otherwise, so ambiguous requirements always route back to a person.

A non-ready environment is never judged as a failure of the feature. The evaluator emits an aggregate of needs-human with empty criteria and a violations note, then tears the environment down, since an unexercised feature was not proven to fail.

No. Reading the builder's tests is an explicit anti-pattern because it re-imports the code's blind spots and breaks the code-blind lane invariant. Integration-tier criteria that require those tests are delegated to the review/merge test floor instead.

When the environment handle carries credentials, derived auth headers such as Authorization Bearer tokens are attached to every request. Tokens are never written to evidence, verdicts, or logs, and the spawner scrubs any leaked secret from persisted output.