open-arena

Operate the Open Arena REST API: start the server and make authenticated requests.

Run the arena evaluation sweep and interpret last_run.tsv and the leaderboard matrix.

Start and operate the autonomous reward-R&D experiment loop (setup, confirm, then run until interrupted).

Add or iterate on a custom reward in src/rewards/ and wire it into the sweep.

Configure an Open Arena sweep: author/edit config.yaml datasets, rewards, metrics, and experiments blocks.

Prepare evaluation datasets: write prepare_data.py, wire loaders in config.yaml, smoke-test that rows load.