DecktraceAutomated capability lab for card-game agents

Decktrace overview · evidence-first card-game agents

One platform.
Many card games.

Decktrace turns each card game into its own deterministic capability lab, then reuses a clear discipline: learn through self-play, compare on balanced starts and seats, and replace the champion only when fresh game outcomes clear the gate. OPTCG and Pokajan are the current implementations, not the limit of the platform.

  • Game outcomes decide
  • Starts and seats stay balanced
  • Claim ceilings stay visible

One platform · game-specific rules and scorecards

  1. 01
    ModelRules and legal actionsversioned per game
  2. 02
    TrainSelf-play policiesno human labels
  3. 03
    MeasureBalanced comparisonsstarts + seats controlled
  4. 04
    RecordDecision + provenanceresult + claim boundary

The method transfers; the meaning of a win does not.
OPTCG measures paired two-seat win rate. Pokajan measures four-seat placement-weighted settlement. Each result stays beside its own plain-language interpretation and simulator boundary.

Current projects

The game result comes first.

Color calls out the conclusions a reader should understand quickly. Technical scope and runtime caveats remain quieter immediately below.

Decktrace: OPTCG · established

The latest champion powers Play.

The latest champion cleared the paired-seat strength and archive checks for the checked-in Imu mirror, and now powers the browser Play surface.

Simulator win rate
53.2%
Against the incumbent in paired-seat games
Conservative floor
50.3%
Promotion measure after uncertainty
  • Paired openings768
  • Archive comparisonsPass
  • Browser Play actorLatest champion
  • Native qualificationSeparate lane
Current decisionLatest champion powers Play

Fresh simulator games and archive checks cleared the automated gate. Native qualification and broad live-game strength remain separate questions.

Confirmed in the checked-in Python simulator for the Imu mirror. Play the latest champion · Inspect the campaign.

OPTCG overview PlayTraining
Decktrace: Pokajan · established

The v2 champion held after five focused follow-ups.

Fresh four-seat confirmation rotated the learner through every seat against three incumbent copies on identical deals.

First-place finishes
39.1%
100 of 256 confirmation games
Top-two finishes
69.1%
177 of 256 confirmation games
  • Average finish2.04 of 4
  • Final coin edge+456 vs mean opponent
  • Archive comparisons6 / 6 pass
  • Invalid actions / timeouts0 / 0
  • Live-client parityNot run
Why the champion is establishedThe gain survived confirmation; later ideas did not

The v2 policy improved first-place rate, top-two rate, average finish, and terminal coins. Five targeted follow-ups then failed fresh game-outcome gates, so retaining v2 is the evidence-backed result.

Confirmed in the versioned Python simulator. Native control, live-client parity, and broad live-game strength are not established.

Pokajan overview PlayTraining

A 21-second platform tour

Watch a card game become measured capability.

The shared evidence discipline, with game-specific rules and scorecards kept visible.

Stage 1 of 7 · Versioned game profileModel: Rules become an executable contract.

Stage 1 of 7Versioned game profile

Model: Rules become an executable contract.

Each project defines its own visible state, legal actions, scoring, and terminal conditions before training begins.

Evidence
OPTCG keeps a fixed two-seat Imu mirror; Pokajan keeps a four-seat hidden-information profile.
Claim ceiling
The platform reuses the method, not one universal rules engine.
Compare the game lanes →
GAME PROFILERULES + SCORE
State
Visible
Actions
Legal
Result
Terminal
game.profile.versioned

Five connected capabilities

The platform at a glance

Open the technical map
Explore the five connected systems
01

Execute

Runs complete games from rules and legal actions that stay specific to each title.

Built
Separate deterministic engines, masked actions, and browser-ready simulator actors
Evidence
The latest champion powers OPTCG browser Play
Current boundary

OPTCG has a native execution/disagreement bridge; Pokajan client parity has not run.

Map the execution lanes
02

Train

Lets policies improve through self-play without human games, labels, or promotion votes.

Built
Game-specific RL modules behind shared-action policies and resumable operator seams
Evidence
OPTCG's latest champion cleared its measured strength and archive gates; Pokajan retained its v2 champion
Current boundary

A completed run is evidence of execution, not automatically evidence of strength.

See what is shared
03

Compare

Controls starting conditions and seats so a policy change—not luck or position—drives the result.

Built
Common openings for two seats; common deals and four-seat rotation for Pokajan
Evidence
The latest champions used balanced comparisons; Pokajan confirmation used 256 games
Current boundary

Win rate and placement-weighted settlement answer different game-specific questions.

Compare the scorecards
04

Confirm

Requires a fresh conservative result and regression checks before a challenger replaces the champion.

Built
Uncertainty-aware lower bounds, invalid-action checks, and archive opponents
Evidence
The latest OPTCG champion cleared its uncertainty and archive gates; Pokajan cleared a +0.867 utility floor
Current boundary

Rules correctness remains a separate prerequisite for any strength claim.

Inspect the evidence gates
05

Improve

Keeps proven gains, diagnoses visible mistakes, and stops ideas that fail game-outcome evidence.

Built
Retained champions, focused diagnostics, bounded probes, and explicit claim ceilings
Evidence
The latest champion is the current OPTCG browser Play opponent
Current boundary

A well-supported stop is a result; more complexity waits for evidence.

Follow the feedback loop

What Decktrace demonstrates

A reusable capability discipline, without pretending every game is the same.

Rules, observations, scores, and runtime boundaries remain game-specific. The inspectable loop around them stays consistent.

See the five platform principles
  • Deterministic simulation and masked legal actions
  • Self-play without human labels or promotion votes
  • Restartable training and source-bound checkpoints
  • Balanced evaluation with uncertainty and archive gates
  • Plain-language results beside explicit claim ceilings