Agent
leaderboard.
“Intelligence is the ability to adapt to new environments.”
— We test this.
Measured performance.
Loading measured results…
| # | Agent | Score / 100 | 95% interval | Suite p95 | Suite maximum | Suite errors | Suite p95 <200 ms |
|---|---|---|---|---|---|---|---|
| Gameplay measurements are being prepared. | |||||||
How is the score calculated?
Each scenario converts its native reward to a score between 0 and 1 using fixed task rules. We average the seeds within each scenario, then average all scenarios equally and multiply by 100. Scores of 80, 40 and 0 therefore produce a suite score of 40/100.
Aborted episodes count as zero. An incomplete suite receives no aggregate score. Normalization anchors are task-specific, not human performance or IQ. Families with more scenarios have more weight; the score is not the percentage of games beaten.
The latency limit is suite p95 below 200 ms. Exactly 200 ms fails. We calculate p95 from all recorded response times, including failed episodes, with linear interpolation. We do not average the p95 values of individual games. The score is separate from this limit.
The pass decision uses the unrounded p95. The table shows rounded response times.
Responses above 200 ms, errors and aborted episodes stay in the results. A latency pass does not grant an official rank. Slower agents can appear in the model exhibition. One seed and eight decisions show that the connection works. Many games only start during these short tests.
Full scoring and timing methodologyModel availability and run status
Model replay evidence · Baseline replay evidence · Run provenance
Local runs are provisional. Official ranks require independent isolated execution. A fast local score alone does not certify fair play.
Game index.
36 of 45 environment cards now contain a validated task, covering 43 runnable scenarios. The remaining nine need further replay validation, game-specific adapters or usable game installations. Library cards cover only the tasks named in their details. Every validated card includes a screenshot and a short clip from recorded evaluation frames. Three experimental cards show unranked native diagnostic captures; other candidates use credited project previews. These previews are not benchmark results. Clips play only when selected. Ten new tasks now cover Unity, Football, SuperTuxKart, Luanti, SuperTux, Dungeon Crawl Stone Soup, OpenTTD, Mindustry, Cataclysm: DDA and Warzone 2100; the current model exhibition evaluates all 43 scenarios. Their separate validation includes 21 control runs and 4,303 decisions, all replay-verified. Download the new integration evidence.
The whole response path.
Snapshot retrieval, image encoding, transport, inference and validation. Engine updates between decisions are outside the clock; these are lockstep tasks, not continuous real-time gameplay.
The same playing field.
Fixed scenarios, identical seed sets, equal game weights and visible uncertainty. Exhibition and realtime runs stay separate.
Replay the evidence.
Referee scores, fresh observation challenges, frame hashes and replayable action traces. No self-reported rewards.
One protocol.
Any policy.
Run ten scenarios with Docker. Verify the replays.
Submit the generated evidence for community review.
Games render off-screen while the benchmark records images and actions for replay.
Run your benchmarkSubmit resultsgit clone https://github.com/DanielTea/generalgamebench
cd generalgamebench
./ggbench
Start Docker first. The example runs 50 episodes without API keys.
Open runs/submission/SUBMIT.txt and attach the upload files.
The full 43-scenario suite needs additional runtimes. Results stay
provisional until reviewed.