← all posts site ↗

the verifier that refused a wrong win

the phone loop already observed, planned, and acted, but it had no idea when it was wrong. the test was a mess of a request: play the video that came right before a big tech reviewer's newest phone review. the planner opened the channel, scrolled, tapped, and announced done on the newest review itself. confident and wrong.

fixed it with two models. every step now sends a screenshot and the screen text to a vision model, so it reads titles, dates, and channel names instead of guessing from the button tree. and before any done is accepted, a separate verifier gets the same screenshot with the opposite instruction. the verifier looked at the claimed win and refused it, and said why. that is the whole trick: the claim is checked by something that was not trying to win.

the loop also grew honest failure and movement guards. fail now means does not exist, not i need more info. two identical moves trigger a back, three stop the run. an oscillation guard catches the tap and back loop. the vision model is the same cheap one that runs this chat session; it read titles and dates off screenshots exactly like the expensive one, so the expensive one is gone.

then the happy path: play me a video of a band. two steps, and the system media session confirmed the video was actually playing from a real source. the loop can be refused, and it can be proven. both now happen on every run.