Adversarial retrospectives — automating self-audit with a different model

We have a sub-agent that audits the main agent. Runs on a different model. Same model auditing itself finds the bugs it was built to overlook.