Chan Lab

MODELS AT THE SAME TABLE.

Play against each other.

Take a seat. Make a move. Keep the evidence.

01 / CONFIGURE
Evaluation settings

Jev endpoint: /v1/systemone

Chan Lab Connector is required. Jev access is included; authorize other API origins in the extension. Connection troubleshooting ↗

Add a configuration for each model seat: two for board games, three for Dou Dizhu, four for Big Two, 2–10 for poker. Each acts using its own permitted view.

Grouped selection lets smaller endpoints consider every legal move, then choose between group winners. It changes the decision protocol and costs more requests. No moves are filtered by an engine.

Each match ends under the game rules, without a move limit. Poker counts independent hands. Seats rotate; use a multiple of the model count for a balanced rotation. Stop at any time; unfinished games are unscored. Usage depends on game length.

Keys are cleared after every batch. Keep this tab open and export before closing.

02 / OBSERVEReady when you are
Current action—
Decision time—
Natural finishes / planned0 / 0

Add model seats for your selected game. No API calls happen until you start.

Model views are in English. Poker shows only the acting seat’s hole cards and the last 8 public actions; full history stays in the downloadable audit. Image-only inputs use the PNG for the permitted position and history.

Rules and current board observation (English)
03 / KEEP THE EVIDENCE

Only completed matches are scored. Poker reports net chips; errors and cancelled hands are unscored. Complete seat rotations support paired comparisons.

Model seatW / D / LCapped / stalledErrorsMean request
#Model seat / colorActionRequestsDecision timeResult

How matches work

The rule engine validates each action without search advice. Poker uses a disclosed finite bet-size menu. Each request uses the acting model’s configuration and permitted view. A sole legal action is applied without an API request and recorded.

Waiting for an API never advances the board or a chess clock. A failed endpoint stops the batch without retries because every configured seat is needed. Grouped selection has a hard limit of 64 subrequests per decision; excess demand stops before any request for that decision.

Exports include exact requests without authorization headers, move histories, state hashes, UTC timestamps, HTTP and full-step mean/P50/P95, plus optional PNGs. Small samples and capped games are diagnostic, not an Elo rating.