One of the 46 Long-Horizon Terminal-Bench tasks, super-mario, asks the agent to actually play the game — issuing moves turn by turn through the terminal for the full 90-minute budget. When the run ends, the hidden verifier replays the agent's own move log in a deterministically-seeded engine and renders the result to video. Nothing here is hand-authored: every clip below is a model's real playthrough, exactly as the grader saw it.

The number on each card is that run's normalized super-mario reward in [0, 1]: full levels cleared plus fractional progress into the current level, divided by the pass ceiling of three levels (clearing through World 1-3 = 1.0). We also show how far each agent actually got. Tellingly, only four models ever cleared World 1-1 — most die on the first stage.