Decktrace Atlas · current platform map
Atlas shows what Decktrace actually reuses across card games and what remains specific to OPTCG or Pokajan. It replaces the former OPTCG-only progress snapshot with a current technical map grounded in the checked-in implementations.
Platform architecture
Each simulator implements its own game rules and legal actions.
The OPTCG native bridge is an execution and disagreement canary. The browser actor is simulator-only.
Each game owns its observations, rewards, and configuration while reusing the masked shared-action approach.
Training volume is execution evidence; OPTCG strength is scoped to the checked-in Imu mirror.
Matching starts and seat rotations reduce variation in policy comparisons.
Balanced sampling cannot compensate for an incorrect game rule.
A challenger must remain better after uncertainty, validity, and regression checks.
A simulator confirmation does not establish native or live-client strength.
Evidence decides whether to promote, retain, diagnose, or stop an idea before adding complexity.
The OPTCG recommendation is simulator-scoped; native qualification and broader strength remain separate.
OPTCG and Pokajan use different game states, action spaces, and evaluation metrics.
The following limits apply to the implementation and its reported results.
Both games use resumable training and automated comparisons. Each defines its own state, actions, reward, matchups, and evaluation metric.
The Play routes package the same simulator-side actors used by the project. Export checks compare browser and Python simulator behavior. Official-client parity requires separate testing.
The trained policy passed simulator evaluation, comparisons with earlier policies, and browser-export checks. These checks do not establish qualification in the original client.
The former Atlas snapshot ends on July 22 and predates Pokajan. It is preserved for historical engineering context, not presented as current platform status.