The model will tell you it worked when it didn’t

Claude Code's report of what it did is a claim, not a result: a feature that counted attempts as submissions, why it's hard to catch, and what I verify now.