DecktraceAutomated capability lab for card-game agents

Decktrace: Pokajan

Four-player self-play,
measured honestly.

A deterministic Pokajan simulator, masked PPO trainer, and automated generation race built beside Decktrace's OPTCG system. Every result below is simulator-scoped.

Watch live training
Simulator scope

Pokajan rules v1

Versioned

Four players draw, declare melds, discard, and resolve simultaneous claims against one exact 100-card table manifest. Hidden hands stay private while public groups, melds, discards, coins, and the remaining-card ledger reach the policy.

Players
4
Starting stake
1,000
Action slots
128
Reward
Final settlement

Claim ceiling: credible strength in the versioned Python simulator. Live-client parity, native control, and broad game strength are not established.

Earlier local evidence

4,096-step PPO

The first trained policy beat both trivial baselines on fresh balanced seeds, resumed exactly from its trusted checkpoint, and remained champion after its isolated Generation 0.

Vs random
+1.403
Lower 95% settlement-margin bound
Vs first legal
+4.852
Lower 95% settlement-margin bound
  • Invalid actionsZero
  • Checkpoint continuationExact
  • Live-client parityNot run
Trusted checkpoint3991872faeb5...

The full SHA-256 remains bound to the retained checkpoint and canonical capability plan.

Automated generations

Generation 0 retained the verified incumbent.

Three candidate hypotheses trained through two cumulative rungs. Fresh confirmation rejected both ranked challengers, so the loop correctly retained the verified champion and advanced to Generation 1 ready.

Generation 0RetainCompleted locally
Automatic decision

Incumbent retained

A positive mean was not enough: no challenger cleared the fresh uncertainty-adjusted lower bound.

Champion continuation
+0.076 mean
-0.219 lower
Fresh candidate
-0.428 mean
-0.796 lower
Controlled play
0 invalid
0 timeouts
Behavior diagnostics

Game outcomes stay separate from explanations.

Settlement decides promotion. Diagnostics make behavior inspectable without turning hand-written strategy judgments into rewards or vetoes.

Recorded now

Controlled match facts

Strength authority
Paired settlement margin
Common manifests, deals, and all four candidate seats
Placement
First-place, top-two, and average finish
Economy
Final coins and coin margin
Safety
Invalid actions and timeouts
Deliberately absent

No invented live story

Live campaign
Runs and publishes through the isolated operator
Detailed play style
Meld, claim, chain, and deal-in summaries are not yet instrumented for public status
Client fidelity
Simulator choices remain versioned until later client comparison

Follow the live campaign and automatic champion decisions.