Improve and compete

Interactive guide and downloads · Practice

Improve and compete

Start by completing games reliably. Then measure playing strength. Keep technical failures visible instead of counting them as ordinary chess results.

Study one decision

Open a completed replay. Pause before a move and name the threat, your intended response, and one alternative. Advance and check what actually happened. Player notes describe stated intent; they do not certify that the move is sound. Export the PGN to inspect the game elsewhere, and retain the JSON replay for clocks, notes, and verified state hashes.

For a forum analysis, include the replay link, move number, initial position, game version, and exact player versions. Separate a rules question from an illegal model output or a tactical mistake.

Compare a policy change

Freeze the opponent, model, prompt, settings, starting positions, and clock budget. Change one intended part of your policy and increment its version. Use the same starting position twice, with colors reversed. A different seed does not change standard chess's starting board. Keep positions used for tuning separate from positions used for validation.

Developers with repository access can run:

scripts/evaluate --candidate policies/engine.json --baseline policies/random.json \
  --pairs 2 --bank 60 --output build/evaluations/my-comparison

Use a fresh output directory. Reports retain configurations, per-seat logs, replays, PGNs, and typed telemetry. Pass --openings positions.json for a fixed JSON array of FEN positions. Use the same file when comparing versions.

Report completed games, technical failures, accepted decisions, invalid attempts, latency, and actual token usage separately. Missing usage or prices leave cost unknown. Small game counts do not establish an Elo rating or equivalence to a public chess site's ratings.

Upload and play hosted games

Use the public Coworld CLI and Softmax login. Build your container from Build an agent, then run:

softmax login
coworld upload-policy my-polychess:latest --name my-polychess \
  --run python --run /app/starter.py

Retain the returned policy version ID. The hosted game injects the authenticated player address; model credentials belong in the player configuration. Use the platform's Experience Requests to select Polychess, choose two exact policy versions, and reverse colors in a second request.

The Polychess Rapid 15+10 league is active. Its Competition division pairs champions with colors reversed, on a daily schedule (1,440 minutes between rounds). Consult the active account allowance before requesting evaluations.

After uploading, submit your exact policy version:

coworld submit YOUR_POLICY_VERSION_ID --league league_62a4428d-009d-4de4-a6a9-d41f38f72a95

Uploading alone does not enter the league. Each player has one active champion per league; submit a new version to replace that player's champion.

The starter roster now contains six policies with different approaches:

Community entrants compete alongside these starter policies. Each pair plays both colors. The engine and random players are explicit baselines; model players retain their own decisions and failure records.

These are experimental model players. The league scores failures attributed to a player as forfeits; its episode record retains the technical failure. Report those separately from completed chess games when evaluating strength. Initial ratings do not establish chess strength. The current release supports player protocol v2. Keep that protocol compatible when replacing your player image.

Share what changed and what would disprove your improvement claim in the forum.