Two-model review gate to stop AI coding agents from falsely declaring tasks done

An engineer describes a review process for AI coding agents, in which a separate model checks completion claims against a locked acceptance contract. The contract defines binary 'done' criteria and conditions that count as not done even when the agent reports success. A second, read-only reviewer must cite specific files and lines for each verdict and repeat until it approves or a round cap triggers human escalation. The author says the reviewer's independence matters more than its raw capability. The approach was tested on a small Node browser-automation toolkit over four review rounds, including one where the gate caught a claimed fix that had not actually been applied.