We Took the Map Away. Gemini 4 Argon Still Beat Brock in 320 Turns.

Gemini 4 Argon holds the PokeBench record with the text map, Brock in 203 turns. This is the same model on the new screenshot-only board. It beat Brock on turn 320 with 1,915 button presses, the fastest of the 14 models that have run it so far. It named itself ASH and cleared dialogue with B from the start. A Bug Catcher's Weedle poisoned Squirtle in Viridian Forest and it blacked out on turn 174, all the way back to Pallet because it had never used a Pokemon Center. It blacked out once more to the Jr. Trainer's Diglett in the gym, came straight back, and took Onix down with two Bubbles. THE RULES (RVB, Raw Vision + Buttons) Each turn the model gets one screenshot and presses up to 6 of the 8 Game Boy buttons. It sees what a player would see and nothing else. Its only memory is its own notes and its recent turns. 1000 turns to beat Brock. Same prompt for every model. All ten milestones are read straight out of the game's memory, so nothing is judged by a human or by another model. MILESTONES, AND THE TURN EACH FIRST FIRED 00:00 Start 00:37 Turn 1 11:33 Left the house .......... turn 15 18:15 Got a starter ........... turn 24 28:28 Reached Route 1 ......... turn 41 34:11 Reached Viridian City ... turn 54 36:35 Got Oak's Parcel ........ turn 60 45:10 Got the Pokedex ......... turn 88 56:32 Entered Viridian Forest . turn 124 1:36:00 Reached Pewter City ..... turn 255 1:40:46 Entered Brock's Gym ..... turn 263 1:57:12 Onix down ............... turn 314 1:58:25 Boulder Badge ........... turn 320 FOR CONTEXT Five of the 14 models beat Brock with only the screenshot. Claude Opus 5.5 was next at 444 turns, then GPT-6 Astra at 478. Every model that beat Brock both ways took 1.6x to 3.1x the turns without the text map. None of the local models got a starter. THE RECEIPTS Leaderboard: https://pokebench.tv/?category=rvb Harness (MIT): https://github.com/VibeCodyH/pokebench Every run publishes its full turn log: the exact message the model saw each turn, its reasoning, what each button press did, token counts, cost, and the commit SHA of the harness that produced it. A bad result goes up the same as a good one. Not affiliated with Nintendo, Game Freak, or The Pokemon Company.