---
type: project
title: "Pokémon TCG AI Battle — Project Memory"
description: "Durable project index for the Pokémon TCG AI Battle agent, completed MLX migration, and local arena/schema overhaul."
tags: [pokemon-tcg, kaggle, mlx, behavioral-cloning, recurrence, elo, schema, replay, dashboard]
timestamp: "2026-08-15T17:24:00-03:00"
status: "aug15-coherence-reconciliation-historical-cohorts-and-sidecar-boundary"
---

# Pokémon TCG AI Battle — Project Memory

## Current state

The live project is `~/workdir/pokemon-tcg`, on `develop` at `20d7d0d` with 144
reachable commits in the verified snapshot. The current trainer and PyTorch
converter enforce strict FP32, and the live agent uses autoregressive
multi-select with per-side scratch memory carried between decisions. The
preserved 14-epoch checkpoint and the user-reported seven-epoch result at
roughly 900–930 ladder rating are historical delivery evidence, not current
acceptance claims. Recurrent TBPTT, Muon/AdamW routing, would-KO integrity and
the local arena/database path remain implemented with the limitations recorded
in the current-state reconciliation. RoPE-ND, MoE, vehicle drafting and Apex
remain documented future or experimental phases; they are not current runtime
modules.

The M3 Pro with 24 GB unified memory is the only development target for this phase. The final arena artifact must be self-contained. The August 14 review cross-checked the live source tree, repository handoff, Antigravity provenance, current Parquet and ZIP files, SQLite, checkpoints, ablation logs and tournament summaries. It did not modify code or repair the database.

The subsequent coherence pass scoped the August 6-7 ablation's “data-limited”
interpretation to that historical cohort because the later Stage 3/4 audit
found a loss-scale and validation-contract defect. Prospective V2 sidecar
references are likewise historical after `a942373`; current auxiliary heads and
would-KO features remain in the primary pipeline, while the sidecar runtime is
not current.

## Canonical reading map

- [[pokemon_tcg_ai_battle]] — public project hub and evidence boundary.
- [[pokemon_tcg_current_state_reconciliation]] — verified current code, database, handoff and Git snapshot.
- [[pokemon_tcg_agent_architecture]] — current architecture and ceiling.
- [[pokemon_tcg_mlx_migration]] — MLX implementation contract and phases.
- [[pokemon_tcg_temporal_learning]] — temporal dataset and recurrence.
- [[pokemon_tcg_tbptt_training_contract]] — exact recurrent training accounting
  for TBPTT, accumulation, scheduler, checkpoints and progress.
- [[pokemon_tcg_training_overhaul_2026_07_29]] — corrected temporal baseline (landing record).
- [[pokemon_tcg_training_pipeline]] — **canonical current training pipeline (start here for anything operational)**.
- [[pokemon_tcg_train_config_reference]] — every field of TrainConfig and every CLI flag of `bc_train_mlx.py`.
- [[pokemon_tcg_kv_cache_hierarchical]] — hot/transient/SSD three-tier row-group cache.
- [[pokemon_tcg_parquet_dataset]] — day-partitioned Parquet format, 90-column schema, row groups.
- [[pokemon_tcg_sqlite_schema_current]] — live schema 2.0.0 snapshot with populated counts, source enum and read-only revalidation.
- [[pokemon_tcg_tournament_system]] — tournament CLI, sweep semantics, JSON reports, intra-suite round-robin.
- [[pokemon_tcg_torch_inference]] — historical FP16 arena delivery reconciled with the current strict-FP32 inference contract.
- [[pokemon_tcg_top_elo_curriculum_filter]] — `--top-elo N` semantics + findings.
- [[pokemon_tcg_glossary]] — canonical vocabulary used across all Pokémon TCG pages.
- [[pokemon_tcg_repository_timeline]] — semantic timeline of the verified 144-commit snapshot, with the older 147-count record preserved as history.
- [[pokemon_tcg_decision_trail]] — chronological + thematic reasoning from working conversations: origin (900→300 Elo regression), sidecar removal, ablation design, feedback rules.
- [[pokemon_tcg_long_horizon_vision]] — staged vision (BC → GRPO offline → 900-Elo world model + scratch zones → MoE), NOT implemented but the trajectory the current design accommodates.
- [[pokemon_tcg_bc_curriculum_ablation]] — historical 10-config days × epochs × top-Elo ablation, per-run tournaments + intra-suite round-robin; its architecture-sanity result is scoped to that cohort and does not supersede the Stage 3/4 incident.
- [[pokemon_tcg_would_ko_prospective_search]] — additive prospective planner.
- [[pokemon_tcg_prospective_v2]] — historical Prospective V2 sidecar,
  group-relative objective, PyTorch runtime reranking and strict-GRPO boundary.
- [[pokemon_tcg_action_coverage]] — deterministic legal-action coverage and
  offline/runtime parity.
- [[pokemon_tcg_data_pipeline]] — replay ZIPs, BC NPY arrays, would-KO
  metadata, historical prospective sidecars, and resumable generation.
- [[pokemon_tcg_daily_corpus_training]] — orchestration and smoke-only boundary.
- [[pokemon_tcg_evaluation_results]] — local tournament observations,
  one-epoch prospective result, and ladder provenance boundaries.
- [[pokemon_tcg_ladder_and_research]] — ladder population and release records.
- [[pokemon_tcg_research_backlog]] — deferred research directions.
- [[pokemon_tcg_local_overhaul]] — local product boundary and decisions.
- [[pokemon_tcg_local_platform_status]] — current local platform implementation
  status.
- [[pokemon_tcg_implementation_specification]] — target services, deterministic
  workflows and acceptance gates reconciled against the live boundary.
- [[pokemon_tcg_schema_evolution]] — populated SQLite debt versus relational v2
  identity, replay and rating design.
- [[pokemon_tcg_arena_future_architecture]] — future application/worker seam
  with the synchronous local baseline preserved.
- [[pokemon_tcg_physical_schema]] — relational schema, constraints and idempotency.
- [[pokemon_tcg_replay_data_model]] — normalized replay and official visualizer handoff.
- [[pokemon_tcg_submissions_and_elo]] — model/deck submissions and rating lineages.
- [[pokemon_tcg_deck_strategy]] — immutable submission deck and local selection loop.
- [[pokemon_tcg_experiments_and_anamnesis]] — generic experiments and temporal observations.
- [[pokemon_tcg_dashboard_and_tournament]] — dashboard, arena and tournament configuration.
- [[pokemon_tcg_aug14_architecture_and_handoff_audit]] — August 14 source-versus-blueprint boundary and Antigravity handoff provenance.
- [[pokemon_tcg_aug14_data_etl_database_audit]] — August 14 read-only ZIP, Parquet, SQLite, day metadata and integrity audit.
- [[pokemon_tcg_aug14_ablations_tournaments]] — August 14 cross-cohort ablation, checkpoint validation and tournament ledger.
- [[pokemon_tcg_curriculum_stage_lineage]] — stage chronology, 871/420/3,600-game cohort separation and checkpoint provenance.
- [[pokemon_tcg_fitalabs_source_set]] — August 14 Codex handoff, analytical thesis and TBPTT/J-space dossier claim boundaries.
- [[pokemon_tcg_docs_corpus_provenance]] — complete 35-file `docs/` inventory, Git provenance timeline, source roles and source-to-page integration map.
- [[pokemon_tcg_empirical_ablation_monograph]] — empirical monograph for the 420-game matrix, deck saliency, FP16 incident and Abelian rating proposal.
- [[pokemon_tcg_data_augmentation_vehicle_draft]] — derived-target versus future synthetic vehicle-draft contract.
- [[pokemon_tcg_handoff_acceptance_reconciliation]] — PROJECT, TEST_INFRA, AGENTS, RFC and acceptance claims reconciled against live evidence.
- [[pokemon_tcg_orchestration_artifact_ledger]] — `.agents` topology, explorer handoffs, milestone states and pending acceptance gates.
- [[pokemon_tcg_stage3_training_failure_postmortem]] — Stage 3 loss-scale, validation mismatch and catastrophic-forgetting analysis.
- [[pokemon_tcg_neural_engine_tokenization]] — as-built tensor streams, set-based tokenization, scratch registers and pointer heads.
- [[pokemon_tcg_ropend_moe_blueprint]] — future RoPEND, stochastic Elo, MoE, 60-card vehicle augmentation and Apex lineage.
- [[pokemon_tcg_data_oracle_etl]] — replay compilation, would-KO/oracle labels, Parquet, zero-trust ETL and synthetic-data provenance.
- [[pokemon_tcg_entity_normalization]] — leaderboard and deck-footprint
  resolution heuristics, saved-probe discrepancy and residual uncertainty.
- [[pokemon_tcg_local_platform_design]] — current arena boundary and future schema/application-worker design.
- [[pokemon_tcg_manuscript_and_empirical_synthesis]] — nine manuscript chapters joined to empirical evidence.
- [[pokemon_tcg_manuscript_chapter_lineage]] — chapter-level Git provenance and
  current evidence reconciliation for the manuscript.
- [[pokemon_tcg_cross_harness_provenance]] — Antigravity report/session provenance and tokenizer/channel adaptation.
- [[pokemon_tcg_harness_session_provenance]] — Portable Antigravity and Claude session locators, 1P/3P classification and evidence boundaries.
- [[pokemon_tcg_elo_pagerank_invariance]] — Abelian Elo, sampled tournament graphs and PageRank analogy.
- [[elo-as-experimental-protocol]] — cross-project comparison of Elo-like evaluation protocols and their non-interoperable scales.
- [[pokemon_tcg_sampling_timezone_identity]] — platform coverage, timezone inference and entity normalization.
- [[pokemon_tcg_metanoia_suite]] — six linked Metanoia specifications and their evidence boundaries.

The documentation lineage also records the previously underrepresented August 8–13 phase: the curriculum sweep orchestrator, GEMINI rule evolution, Stage 3 and auxiliary-loss changes, manuscript landing, Abelian Elo, schema and ETL audits, sampling/timezone analysis and the first RoPEND/MoE blueprint. See [[pokemon_tcg_repository_timeline]] for the commit-level sequence.

## Source authority

Operational state comes from the live repository and artifacts. The originating conversation supplies explicit product decisions and hypotheses. Code evidence takes precedence for implementation claims; Antigravity reports are provenance records; user-reported training results remain labeled as such until corroborated by repository artifacts. The August 14 audit pages preserve unresolved discrepancies instead of silently selecting a convenient number.

## Durable boundary

The current live implementation does not include Mamba, Hope, RoPE-ND, Energy-Based Transformer, TRM, J-Lens, strategic MoE, PPO/GRPO/GSPO, world models in inference, symbolic regression, custom Metal or distributed Mac training. The August 14 repository documents preserve RoPE-ND, MoE, vehicle-draft and later phases as planned or experimental work; those ideas remain in the backlog and handoff map in [[pokemon_tcg_ladder_and_research]], [[pokemon_tcg_aug14_architecture_and_handoff_audit]], [[pokemon_tcg_ropend_moe_blueprint]] and [[pokemon_tcg_docs_corpus_provenance]].

The August 15 documentation pass also records the Stage 3 training incident as a first-class research result. A validation score is not treated as sufficient evidence of policy health when the split, objective denominator and tournament cohort are misaligned; see [[pokemon_tcg_stage3_training_failure_postmortem]].

The current-state anchor is [[pokemon_tcg_current_state_reconciliation]]. It
must be read together with the dated experiment pages: a newer source
snapshot can correct a runtime claim without invalidating the empirical cohort
that produced an older result.