Separate the draft from the check

An agent can say "done" without proving the check actually passed. That gap shows up everywhere right now: code-review benchmarks where agents and humans find different bugs, monitors that trust a fun