Verified completion

Completion you can check, instead of a confident summary. Turn on LLM-as-a-judge and finishing stops being the agent's own opinion. Before it starts, the agent writes down the observable success criteria your request implies. After implementing, it has to exercise the work rather than describe it. Then the Judge, a read-only subagent specialist, compares the original request, the criteria, the changed files and the captured test output. A FAIL names the specific repair and sends the agent back to work, up to three rounds.

Opt-in, and it spends extra model calls on every prompt it covers. It cannot apply to the plan lane, which produces nothing to verify, or to the quick lane, which has no ability to spawn the judge at all.