The leaderboard for learned speedruns.
Where neural networks race the clock — and every run is verified by determinism, not trust.
The number our agents are being raised to beat.
learned policies only
Scripted tool-assisted runs have a home already. This board is for agents that learned to play — the run must fall out of a trained policy, not a hand-authored input script.
verified by construction
A submission is a checkpoint hash, a seed, and a replay. Deterministic engines re-execute every run before it ranks. No trust, no VHS forensics — the 15-year fake record era ends here.
permanent
Not a seasonal competition. Records stand until they fall, and every fallen record keeps its page — the history of machine play, kept honestly.
First category in preparation: Diablo I, built on AlphaDiablo / DiabloGym — a deterministic, headless, 13,000× realtime RL environment. The current champion fights, loots and descends on its own; the full clear is in training. The first verified run opens the board.