Chan Lab

DECISIONS, ONE MOVE AT A TIME.

Let your model play.

Same seeds. Separate boards. Every move recorded.

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 ↗

For a comparison, add each API separately, including its input mode. Keys stay in memory and are cleared after the run.

Run budget per APIstart small ↙

Episodes × maximum actions. Games stop early at death or victory.

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

02 / OBSERVEReady when you are
Score · current episode—
Mean API round trip—
Completed episodes0 / 0

Choose a game. No API calls happen until you start.

The board is always in English, including image inputs. An image saved as evidence is only sent to the model in Image or Text + image mode.

Rules and exact state (English)
03 / KEEP THE EVIDENCE

Downloads include seeds, states, actions, UTC timestamps and timing distributions. Failed or unfinished episodes stay separate from completed scores.

#Episode / stepActionScoreResultAPI time

What this measures

Each API receives an independent environment with the same initial seed. APIs take turns; the first API rotates every step. Waiting for a response never moves the game. Different actions create different trajectories, so later boards may diverge.

2048 ends at a 2048 tile, no legal move, or the action limit. Snake uses a 10 × 10 grid, one cell per action, with no starvation rule. Allowed Snake actions exclude reversal but may still collide. Short runs are diagnostic; compare completed episodes on common seeds, not general model intelligence.

The model receives the current state and rules, never the random seed or future food. A network, CORS, timeout or invalid action stops that API; other APIs continue. Exports contain the exact model request without authorization headers.