Pokémon TCG — Manuscript Chapter Lineage

Role and Git origin

The nine files under docs/manuscript/ and the integrated docs/Pokemon_TCG_AI_Monograph.md were introduced in project commit d9c67a7 on 2026-08-12. They form a narrative explanation of the project rather than a second implementation source. The empirical companion docs/empirical_ablation_monograph.md was added in the later specification wave at cec2eab on 2026-08-14.

The manuscript is preserved because it records the questions that connected engineering choices to research hypotheses. Its claims are read through the current code, artifacts, database and tournament cohort pages. A compelling sentence in a chapter is not by itself a code-validated result.

Narrative arc

flowchart LR
    B[Behavioral cloning cold start] --> D[Replay cleaning and typed data]
    D --> T[Set tokenization and epistemic features]
    T --> O[Engine-backed would-KO oracle]
    O --> P[Pointer action surface]
    P --> M[Scratch memory and TBPTT]
    M --> E[Curriculum and auxiliary objectives]
    E --> V[Tournaments, Elo and inference]
    V --> F[GRPO, routed and synthetic-data futures]

This is a research progression, not a monotonic performance proof. The Stage 3/4 incident shows that additional objectives can produce an optimization and validation defect even when an isolated validation metric appears plausible.

Chapter-by-chapter reconciliation

Chapter	Narrative contribution	Evidence-bound current reading
01 — Introduction and teleology	Behavioral cloning injects a valid-action prior into a sparse-reward POMDP; pure imitation may inherit oracle entropy; GRPO is the proposed later transition.	BC is the implemented training family. The ceiling and GRPO claims are research framing; GRPO/self-play is not the current runtime. See pokemon_tcg_training_pipeline and pokemon_tcg_research_backlog.
02 — Data pipeline and KV cache	Replays are semantically filtered, action/state alignment is repaired, rewards telescope and Parquet row groups are cached for TBPTT.	The compiler, Parquet catalog and hierarchical cache are live concepts. Counts, drop rates, cache behavior and reward denominators require the ETL and Stage 3 evidence pages; source prose does not override their audits.
03 — Tokenizer and epistemology	The model uses typed zones and unit aggregation instead of ordinary sequence position, with explicit uncertainty signals.	Set-based streams and scratch/memory interfaces are current. “Abolition of RoPE” describes the as-built endpoint; RoPE-ND remains a future architecture, not a contradiction.
04 — Engine simulation / would_ko	A bounded engine simulation generates prospective lethality, prize and terminal targets for legal attacks.	The target contract and engine-backed labels are documented. The oracle is a label generator, not an observed outcome, and its sparse denominators must be separated from policy validation.
05 — Action encoder and pointer heads	Variable legal-action sets are represented by source/target/action features, compact option buckets and a separate submit head.	This matches the current architecture and runtime boundary, subject to exact source/configuration for a checkpoint. Multi-select behavior is current autoregressive selection, not a static top-k claim.
06 — Scratch registers and anomalies	TBPTT carries scratch state; the Stage 3 loss spike is narrated as a shock that the scratch registers absorbed.	TBPTT and recurrent memory are source-backed. The “scratch absorbed the shock” explanation is a hypothesis from the historical report; the validated incident is the auxiliary-loss scale and validation-contract defect. See pokemon_tcg_stage3_training_failure_postmortem.
07 — Empirical results and Elo	Logistic/MD10/Abelian rating ideas interpret tournament results and an older 871-game matrix.	Rating code and stored observations exist, but each number belongs to a declared cohort, deck, seed and reset context. The PageRank connection is a formal analogy, not a verified implementation-level isomorphism.
08 — Curriculum, optimizer and DevOps	Curriculum V1, dual optimizer, learning-rate stages, telemetry and future Stage 5 are presented as one operating arc.	Trainer and telemetry pieces are real; stage labels and historical scripts must be checked against pokemon_tcg_curriculum_stage_lineage. Stage 5/Apex material remains future or experimental.
09 — Tournament, database and inference	Deck sweeps, relational match evidence and an MLX-to-PyTorch deployment split close the loop.	Tournament persistence and local deck sweeps are current. The old FP16 deployment narrative is historical; the current conversion and inference contract is strict FP32 despite a legacy checkpoint filename.

Claims that changed meaning over time

Three chapter-level statements need explicit temporal qualification:

1. Architecture versus data: the Aug 6-7 ten-run suite did not expose a broad encoding, cache or collapse failure within that cohort. Later Stage 3/4 analysis found a real auxiliary-loss and validation defect. The former cannot be generalized into “architecture is not the bottleneck.”
2. Performance peak versus mechanism: the Stage 3 deck-conditioned peak is an observation in a historical evaluation cohort. It does not prove that a loss spike improved the policy or that scratch registers causally absorbed a gradient shock.
3. Deployment precision: the manuscript's FP16 description records the earlier conversion crisis. The current live boundary is FP32 validation in the PyTorch converter and inference path.

These are not editorial deletions. They are provenance-preserving updates that keep the narrative while routing current claims to current evidence.

Evidence graph for manuscript use

Narrative layer	Primary project source	Current Wikifita evidence page
BC and curriculum	rl/policy_mlx.py, scripts/bc/bc_train_mlx.py, stage artifacts	pokemon_tcg_training_pipeline, pokemon_tcg_curriculum_stage_lineage
Replay and oracle	dataset compiler, engine probes, Parquet manifests	pokemon_tcg_data_oracle_etl, pokemon_tcg_parquet_dataset
Tensor and memory	rl/policy_mlx.py, inference adapter, trainer	pokemon_tcg_neural_engine_tokenization, pokemon_tcg_tbptt_training_contract
Failure history	training diffs, logs, ablation artifacts	pokemon_tcg_stage3_training_failure_postmortem, pokemon_tcg_empirical_ablation_monograph
Tournaments and ratings	results.db, tournament scripts and cohort reports	pokemon_tcg_tournament_system, pokemon_tcg_aug14_ablations_tournaments, pokemon_tcg_elo_pagerank_invariance
Deployment	rl/policy_infer_torch.py, converter tests and checkpoint records	pokemon_tcg_torch_inference
Future research	architecture, augmentation, RFC and backlog docs	pokemon_tcg_ropend_moe_blueprint, pokemon_tcg_data_augmentation_vehicle_draft, pokemon_tcg_research_backlog

Editorial rule for future chapter updates

When a chapter is changed, the Wikifita update should record four separate facts:

• the introducing or editing project commit;
• whether the statement is current implementation, empirical observation, inference, hypothesis or future design;
• the artifact or code path that validates or limits it;
• the evaluation cohort and denominator when a number is involved.

This keeps the manuscript alive as research history without allowing it to silently become a derivative source for reconstructing the project.

Source anchors

• docs/manuscript/01_introduction_and_teleology.md through 09_tournament_orchestrator_and_inference.md, introduced at d9c67a7.
• docs/Pokemon_TCG_AI_Monograph.md, introduced at d9c67a7.
• docs/empirical_ablation_monograph.md, introduced at cec2eab.
• pokemon_tcg_docs_corpus_provenance — complete source inventory and Git timeline.

Cross-references

• pokemon_tcg_manuscript_and_empirical_synthesis — existing manuscript hub.
• pokemon_tcg_neural_engine_tokenization — tensor and tokenizer layer.
• pokemon_tcg_data_oracle_etl — compilation and target provenance.
• pokemon_tcg_stage3_training_failure_postmortem — failure chapter expanded.
• pokemon_tcg_aug14_ablations_tournaments — denominator-separated results.
• pokemon_tcg_torch_inference — current FP32 deployment boundary.
• pokemon_tcg_repository_timeline — chapter introduction and later edits.