The Check Stays Failed

Give an AI assistant an account brief to prepare, with a source required for every customer commitment. Then give it a record with one source missing.

Watch the second draft. Does it flag the gap for the account owner, or quietly drop the commitment so the brief passes? A cleaner page can hide unfinished work.

Anthropic’s software-verification research used a separate checker to assess Claude’s proofs across 239 test entry points. An account brief won’t get mathematical certainty. It can still have a source requirement the assistant cannot remove.

I’d let the assistant keep looking, and let it say it’s stuck. The account owner can decide what to do with the missing evidence. Until then, the check stays failed. Even if the brief is beautifully written.