Claims checking

Models sometimes say they did something they didn't. Crewdesk checks.

Before a reply reaches you, a verifier compares what it claims with what the tools actually returned in that turn:

• Artefact claims: "I made a page", "I saved the file", "here is the report".
• Outcome claims: "I sent the e-mail", "the automation ran", "the deploy succeeded".
• Behaviour claims: "this now works", "the tests pass", "the page looks right on mobile", which need the check, the run or the screenshot that shows it.

If a claim doesn't match, the reply goes back to the agent once, with what's wrong, and the agent continues from where it stopped: it does the missing step or corrects the reply. Writes are reported in their own terms (a sent e-mail as "sent", an updated record as "updated"), not in a generic "done".

The agent is also told to investigate before acting: look things up before writing, and route each question to the tool that can answer it.