Finally I got access to @typesafeai (thank you guys!)

Finally I got access to @typesafeai (thank you guys!) Jev + emulator lookahead cleared Super Mario Bros. 1-1 only 0.04$. In this setup, Jev sees no pixels: I parsed RAM into structured state, simulated controller macros, filtered predicted deaths, and let Jev choose. Recorded the run; resumed from the same state after an API outage. The tradeoff: RAM gives precise, auditable observations, but my adapter is game-specific. Bypassing vision makes this experiment easier; it doesn’t solve perception for general game agents. The potential is Jev as a tactical decision layer over structured observations and predicted outcomes. Caveat: 1-1 is easy for RAM + search. Next: Jev vs a cheap heuristic on identical simulations. Does the model actually earn its place?