Pokémon TCG: 8,400 game records across 7 opponent implementations

What was built

• 4-arm Lucario matchup experiment: tests 4 policy arms: Baseline, Card only (replacing 1 Carmine with Xerosic's Machinations to reduce a large opposing hand to 3 cards at resolution), Rule only (preferring visible Abra or Kadabra targets) and Combined.

  Source: docs/STRATEGY-WRITEUP.md and docs/EVIDENCE-MAP.md Status: Public experimental design and analysis at commit aad907a.

• 256-unit recurrent v2 learner interface and 7 selection checks: encodes acting-player visible state, carries a 256-unit recurrent memory across decisions, scores legal options, and steps through multi-pick prompts with a STOP action across 7 component selection checks.

  Source: docs/learner-evidence/README.md, action_decoder.py.excerpt.txt and proofs/selection/README.md Status: Public source excerpts and runnable CPU component checks at commit aad907a.

• Research reader repair tested on a synthetic 4/4 wins report: preserves introductory text before section headings so a top-of-file retraction on a synthetic 4/4 wins report survives retrieval and search.

  Source: FACTORY-IN-ACTION.md and 13-memory-reader-regression.txt Status: Published source and synthetic regression excerpt at commit aad907a.

Exact results

• 8,400 game records across 7 opponent implementations: records 600 baseline games and 600 counter games against each of 7 opponent implementations, scoring 1 for a win, 0.5 for a draw and 0 for a loss.

  Source: workflow/research/2026-09-03-strike3-rows.csv and METHODS.md Status: Published CSV game records at commit aad907a.

• +20.42 and +18.75 percentage points against the 2 tested Alakazam implementations: the Combined arm lifted mean game score by +20.42 and +18.75 percentage points against the 2 tested Alakazam implementations.

  Source: docs/STRATEGY-WRITEUP.md, 2026-09-03-strike3-rows.csv and 02-matchup-results.png Status: Published CSV rows and summary table at commit aad907a.

• -0.47 percentage points mean change across the other 5 opponents: across the remaining 5 opponent implementations, the mean change was -0.47 percentage points with a 95% interval of [-2.64, +1.71] and individual opponent changes from -3.08 to +0.67 percentage points.

  Source: docs/STRATEGY-WRITEUP.md and 2026-09-03-strike3-rows.csv Status: Published CSV rows and summary table at commit aad907a.

• 23 wins in 1,200 decided games for the earlier August imitation checkpoint: the earlier flat-action imitation checkpoint won 23 of 1,200 decided games against the reference policy, which motivated the v2 legal-option interface.

  Source: docs/EVIDENCE-MAP.md and docs/learner-evidence/README.md Status: Historical evaluation of the August checkpoint; v2 full-game win rate is not measured.

October 30, 2026: 5,600-game closed-loop action scorer confirmation & ESS = 48.06 diagnosis

In our local PTCG-Learning-Lab workspace, we trained Ridge regression action scorers on 511 eligible first-search development interventions (512 games total) and evaluated them in closed-loop play across three fresh confirmation studies totaling 5,600 games and 1,002,605 verified action records. All three studies proved that the learned weights changed live search choices (266, 812, and 868 changed choices), and all three failed the pre-registered promotion gate requiring a 95% bootstrap confidence interval strictly above 0.0.

• 58 score-alias groups across 236 of 511 development rows: in the 88-feature representation, 58 groups of distinct game states produced identical feature vectors despite different visible hands (45 groups also differed in boards), such as dev-0010 vs dev-0145 on target card 647.

  Source: PTCG-Learning-Lab/checks/tcg-scorer-diagnosis-v1/diagnosis.json (sha256: ad5cebda...d50f) Status: Verified local diagnosis; resolved at the feature level in tcg-relational-v1.

• 58 / 511 exact-parent matches (11.35%) and parent_ess = 48.06 under uniform exploration: across the 511 eligible development games, the player faced a mean menu of 10.865 legal options, so uniform 1/K exploration matched the exact parent choice only 58 times (14/128, 15/127, 10/128, and 19/128 across the 4 opponent/seat strata), yielding parent_ess = 48.065 and max inverse-propensity weight 27.0 against a 100-observation direct-comparison screen.

  Source: Machine learning & AI learning laboratory and learning-evidence-record.json Status: Published support diagnosis; competitive player improvement remains unproven.

Verify in one minute

Open workflow/research/2026-09-03-strike3-rows.csv and docs/STRATEGY-WRITEUP.md to inspect the 8,400 game rows and the +20.42 and +18.75 percentage point Alakazam gains, or clone thisisntjon/poketcg-research at commit aad907a:

git clone https://github.com/thisisntjon/poketcg-research
cd poketcg-research
git checkout aad907a5b4f9f20de5f591ec49aa2b7ed9c46ea2

Scope of these checks

• Complete-game win rate for the v2 learner: the 7 selection checks verify the v2 legal-option decoder with fixed scores on CPU; complete-game competitive strength of the v2 learner has not been measured.

  Source: docs/learner-evidence/README.md, proofs/selection/README.md and Pokémon TCG case study Status: Component interface checks only; no v2 full-game win rate measured.

• Non-redistributed assets: trained model weights and 1 development host's raw rows are not redistributed in the public repository.

  Source: docs/EVIDENCE-MAP.md and Pokémon TCG case study Status: Documented repository boundary.