How to reproduce

1. Pick a harness, an engine and a model. Start the engine with the model loaded, and nothing else on the GPU.
2. Make an empty folder and start the harness in it with the exact text of challenge/v1/PROMPT.md as the only message. Turn off any time limit in the harness.
3. Stop the run when the harness says it's done, or by the stop rule: no tool call for 30 minutes, or the same failing action 4 times in 30 minutes with no file change.
4. Score it: cd checklist && npm install && node check.mjs /path/to/build --out /path/to/out. It needs Node 20+ and a Chromium (set CHECK_CHROME if it isn't at /usr/bin/chromium). It writes checklist.json, screenshots and a clip.
5. Compare your checklist.json with the matching result's data.

The checklist in this repository is the same code that scored every result. Its version is printed on each result page.