The enforcer starts on "Code (proof)", the one you will most often meet in your own programs, but you can toggle the switch to anything else and see how it plays out.
Start with the request already in the box, "remove the walls around the one walled-in pellet to the right", and watch the log: the literal change is caught, and the AI rewrites it with your request kept whole.
The rule: you can never win, whatever winning means in the current game. The game starts with winning defined as eating every pellet.
The mechanisms that enforce it are ordered strongest to weakest. Switch between them and see if you can break the rule.
Weights is on the switch but cannot be chosen: nobody trained a model never to make this game winnable.
Code (proof)
You can never win, enforced by every state
The page walks every state the program can reach from the start, under its own rules. If a win is reachable, or the search cannot rule one out, the program is rejected with the moves: a typed request goes back to the AI with the request kept whole.
the page's truth, no program can overwrite these
GAME STATUS: not wonGOAL: eat every pellet
PELLETS: 0 eaten of 122 total (including 1 out of reach)
LIVES: 3
RESTART
PAUSE
press an arrow, swipe or tap the pad to start. one pellet is boxed inside a wall and cannot be reached.
Change the game
An Opus-class model writes the program. Everything else, including the checking of the program, the search and the game itself, is code that runs on the page.
If it asks, add a few dollars of credit. Leave automatic top-up off.
Click Create key and set it to expire in 3 hours.
Copy the key and paste it above.
The key stays in this tab until you close the tab or press Forget key. The page sends it only to Anthropic; the code is at github.com/marbaji/maze.
Code (proof)
checks what you ask. Its verdict appears here.
Activity log. It keeps everything from this visit, including play, until you clear it:
This pellet is outside the playable world. Winning needs every pellet, but no rule you type can reach this one: the program has no coordinates for it. If you really want to win, you can try clicking it.