Keep the failure evidence.

An agent reporting completion does not establish that a task met its acceptance criteria. The result needs an inspectable check.

I built Evalgate to connect Markdown work contracts to executable checks, record their results, and leave failed checks unresolved for inspection and an explicit retry.

Define a verifier alongside the task and use its result to update completion. Shell and HTTP checks test explicit conditions; model-based judgment carries different uncertainty.

A passing check proves only the property it tested. Retries are explicit and the declared retry count is advisory. JSON checks cover a limited shape; file-pattern checks are not a semantic-diff engine.

1. Constraint: An agent report is not acceptance evidence.
2. Decision: Run an explicit verifier and retain its output.
3. Result: Passing contracts update; failures remain pending for inspection and explicit retry.

Source-verified implementation pattern; sample outcomes are illustrative. A pass covers only the checked property. Retry counts are advisory; model judgments differ from deterministic checks.

Real local sample execution · two passed · one deliberate failure · synthetic checks.

The CLI parses pending contracts, runs their verifiers, prints failure output and updates the Markdown using the results. In the sample run, the failed source-evidence contract remains pending.

Sample results are demonstration evidence, not product performance metrics. Passing checks do not prove unchecked properties.