Docs
Documentation
How the benchmark is built and run: the harness, the agent and what it is sent, scoring, and how to set it up yourself.
Start with How it works, then read whichever part you need.
- 01How It WorksThe pieces, and one run from start to finish
- 02The BenchmarkWhat is measured, scoring, running episodes and comparing runs
- 03The AgentHow the model plays: turns, rules, memory, failure handling
- 04ContextExactly what the model is sent, and how the context stays within budget
- 05Tool ReferenceEvery tool: parameters, behaviour and results
- 06The HarnessUsing the dashboard, model profiles, run files, videos, safety boundaries
- 07Publishing RunsPackaging runs, scrubbing them and uploading them to the Hugging Face dataset
- 08SetupInstalling the game machine and the host
- 09The ReaderHow game state is read, what it contains and its limits
- 10StatusWhat is verified, what is not, and known issues
The prompts the model receives are in prompt/: the controls guide,
the game reference and the benchmark rules.
What is planned next is on the roadmap.
These pages are docs/ in the benchmark’s open-source repository, at commit cad91f7 (13 October 2026). View them on GitHub