A clean AI summary can still be wrong, and you send it anyway. Frank's adversarial audit skill checks one against its transcript. Verdict: redo it.
