A robot. A snack. One decision at a time.

Jev chooses the robot's next move toward the food. The server works out which orthogonal moves are legal and asks for a typed Choice between them, with what the robot has learned on this map attached to each option. When only one move is legal the code takes it without a call ("Forced move • code"). If the robot starts circling, a deterministic local loop recovery commits to the nearest unexplored cell and walks there through cells it has already seen — those steps are labelled and carry no model numbers. After a successful run the route it actually walked is remembered and can be replayed.

Memory is per map and lives in this browser: it survives Reset and a page refresh, and "Forget learning" returns the robot to cold behaviour. Each attempt is capped at 256 moves — enough to explore every cell of an 8×8 map several times over — and the cap is checked before anything is requested, so a retry cannot spend past it. The model's own weights never change; only what it is told and when recovery takes over.