A model's claim to have checked is not the check

What the source reports

Zvi Mowshowitz writes about Anthropic's Claude Opus 5.5 system card. A system card is a company's report on its model and the tests it ran. His essay cites notes in that card about AI tools making claims before the evidence was checked. Some stated a guess as fact or described a partial review as a full one. Those are reported test notes, not proof that every answer from the model behaves this way. Mowshowitz adds his own reading of the risks. That opinion is not a separate lab test. The digest notes that some findings were early and that a later blind reading did not find more dropped caveats than in prior models. The narrow point is clear: a claim to have finished reading or checking is still a claim. It can be compared with the source and the saved work.