Marcus Wong Software Engineer, AI Systems
Automated capability lab
Decktrace trains several card-game agents in parallel, eliminates weak candidates, and promotes a new champion only after fresh paired evidence clears uncertainty and archive checks. The entire loop runs without human games, labels, or promotion decisions.
One generation · one controlled capability race
fresh + champion lineages256,000 steps each2,048,000-step horizon512+ paired openingsFirst verified promotion: 97.66% point estimate · 96.29% adjusted lower bound.
512 paired openings · archive pass · zero invalid actions.
A 21-second guided system tour
Real retained artifacts from across the system—not one implied end-to-end run.
Stage 1 of 7Retained native trace
The native game records the player-visible action, actor, turn, and trace position before any downstream interpretation.
event.activate_leaderThe engineering challenge
Challenge: Long training runs must survive disposable machines, dependency drift, interruption, and publication failures.
Response: One zero-touch operator restores validated state, owns locking and dependencies, checkpoints complete waves, publishes status, and pauses safely.
Challenge: Every training idea can consume millions of steps before revealing that it is weak.
Response: Five explicit arms train in parallel and narrow from five to three to two on common-seed screens.
Challenge: A top training score can still be noise, seat bias, or regression against older opponents.
Response: Fresh paired openings, adjusted lower bounds, invalid-action checks, and a separate archive gate decide promotion automatically.
First measured capability gain
The counterfactual explorer emerged from the automated multi-arm race.
512 paired openings put challenger and incumbent in both seats across 1,024 completed games.
The challenger posted a 97.66% point estimate against the initial champion.
The multiplicity-adjusted lower bound was 96.29%, above the predeclared promotion threshold.
Archive regression passed, invalid actions stayed at zero, and the promotion happened automatically.
Concrete proof
Each row names what the evidence permits—and where it stops.
Restartable, parallel training has advanced multiple generations with retained checkpoints and live status publication.
97.66% point estimate; 96.29% adjusted lower bound across 512 paired openings; archive pass; zero invalid actions.
The promotion evidence comes from the deterministic simulator; the native client remains a later qualification canary.
The confirmed result is intentionally limited to the checked-in Imu mirror matchup.
Five connected capabilities
Runs complete, reproducible games fast enough for large self-play campaigns.
Rules and strength remain scoped to the fixed Imu mirror setup.
Explores several learning strategies without turning every idea into a full run.
Rung scores screen candidates; they do not prove strength.
Ranks candidates on the same openings with both seats represented.
The comparison establishes only fixed-matchup simulator performance.
Requires uncertainty-aware evidence before a challenger can replace the champion.
Native qualification remains separate from simulator promotion evidence.
Carries the winning policy forward while retaining older champions as regression pressure.
A confirmed matchup gain is not broad strength across arbitrary decks.
Featured engineering artifact
The retained sample records 134 actions from the real game runtime through a turn-five terminal state. Its source identity and configuration are bound to the artifact; parity evidence compares separately retained traces and does not imply that this one file proves the full lifecycle.
What the campaign tests next
What this project demonstrates
Designed and built by Marcus Wong across simulation, agent interfaces, evaluation, and experiment infrastructure.