---
type: reference
title: "Pokémon TCG — Manuscript and Empirical Synthesis"
description: "Reading map for the nine manuscript chapters and their connection to current code, ablations, tournaments, ratings and unresolved hypotheses."
tags: [pokemon-tcg, manuscript, monograph, ablation, tournament, elo, research-history]
timestamp: "2026-08-15T17:42:30-03:00"
---

# Pokémon TCG — Manuscript and Empirical Synthesis

## Why the manuscript is a separate layer

The nine `docs/manuscript/` chapters and the integrated monograph are not simply README pages. They form the project's explanatory narrative: why behavioral cloning was used as a cold start, how replay data became typed tokens, why engine-backed oracles were introduced, how temporal memory and auxiliary supervision were justified, and how tournament outcomes changed the design. The separate `docs/empirical_ablation_monograph.md` is the quantitative companion to this narrative, not a duplicate of it.

The narrative is valuable, but it is derivative relative to code, artifacts and database reads. When a manuscript statement conflicts with a direct run log or current source, Wikifita preserves the manuscript as a historical explanation and updates the current-state page with the conflict.

## Chapter map

| Chapter | Core question | Evidence or boundary |
|---|---|---|
| 01 — Introduction and teleology | Why begin with behavioral cloning and what ceiling does imitation impose? | Design rationale; BC ceiling needs empirical evaluation |
| 02 — Data pipeline and KV cache | How do raw replays become a streaming training corpus? | Parquet/cache contracts and ETL audit |
| 03 — Tokenizer and epistemology | What does a typed set representation expose or hide? | Neural spec and feature-availability contract |
| 04 — Engine simulation / would-KO | How can the rules engine provide prospective labels? | Oracle provenance, bounded simulation and sparse targets |
| 05 — Action encoder / pointer heads | How is a variable legal-action surface scored? | Option buckets, masks and multi-select runtime |
| 06 — Scratch registers / anomalies | How does TBPTT retain temporal context and what broke? | Stage 3 postmortem and loss-scale history |
| 07 — Empirical results / Elo | How should tournament outcomes and ratings be interpreted? | Cohort-separated ledger and Abelian formalization |
| 08 — Curriculum / DevOps | How are stages, optimizers, checkpoints and telemetry operated? | Trainer source, stage logs and handoff records |
| 09 — Tournament orchestrator / inference | How are artifacts evaluated and packaged? | SQLite tournament records, deck effects and runtime boundary |

## Integrated research arc

The documents describe a sequence of increasingly explicit constraints:

```text
behavioral cloning
    -> typed replay and action surface
    -> engine-backed prospective labels
    -> recurrent scratch/TBPTT
    -> curriculum and auxiliary objectives
    -> FP32 and evaluation-cohort separation
    -> rating-aware, routed and synthetic-data blueprints
```

The sequence is not a monotonic proof of improvement. The Stage 3/4 postmortem shows that adding prospective targets can create a multi-task optimization failure even when validation metrics remain plausible. The empirical lesson is that every architectural addition needs an objective ledger and a tournament cohort with a declared denominator.

## Empirical anchors

The manuscript's empirical chapters connect to several distinct result families:

- full round-robin records: Stage 1 28.47%, Stage 2 28.93%, Stage 3 epoch 31 30.42%, Stage 3 epoch 32 30.19%, teacher 67.16%;
- 20-game benchmark: Stage 1 14.17%, Stage 2 14.31%, Stage 3 14.03%, Stage 4 14.72%, teacher 43.89%;
- comparable 420-game FP32 matrix: Stage 1 14.29%, Stage 2 15.24%, Stage 3 13.81%, Stage 4 17.14%;
- Stage 4 deck-conditioned slices: Deck #633 at 27.9% and Deck #251 at 12.9% over 140 games each;
- older ten-configuration curriculum suite: validation and public/peer tournament values are preserved as a separate cohort.

The numbers should be read as a matrix of observations, not one leaderboard. Opponent labels, deck selections, seeds, game counts, reset history and source cohorts differ. The current database preserves 128 tournament summaries, but it does not prove that all underlying match rows survived every reset or rebuild.

## Theoretical bridge to Elo and PageRank

The manuscript formalizes an invariant rating view in which additive translation of all latent strengths should not change pairwise outcomes. The PageRank document then compares this tournament graph behavior to a Markov-style stationary ranking over Wikifita/project relations. This is a useful structural analogy and a possible research bridge, not an established isomorphism in the current system.

The correct claim is therefore:

```text
the project has a formal Abelian rating proposal and a PageRank analogy;
the current tournament database and code provide evidence for the data surface;
the claimed spectral equivalence remains a research result to validate.
```

## Manuscript claims that require care

- “Scratch absorbed the shock” is an Antigravity explanation, not a direct gradient attribution.
- A high validation accuracy is not a proof of arena strength.
- FP32 restored stability in the documented inference collapse, but it did not by itself establish better data or loss semantics.
- The future RoPEND/MoE and vehicle-draft material is a research trajectory, not shipped runtime.
- Metanoia language describes research protocol and philosophy; it is not mechanistic evidence about model cognition.

## Editorial use

The manuscript layer should be retained because it records how the project connected implementation decisions to research questions. New Wikifita pages should link to the chapter that introduced a concept and to the current evidence page that validates, narrows or disputes it. This avoids two common losses:

1. reducing a rich research history to a current-state checklist;
2. promoting a compelling narrative sentence into a code-validated fact.

The chapter-level Git origin, claim-by-claim status and narrative-to-evidence
routing are maintained in [[pokemon_tcg_manuscript_chapter_lineage]]. This
keeps this page as the synthesis hub while giving each chapter a recoverable
provenance path.

## Primary sources

- `docs/Pokemon_TCG_AI_Monograph.md` — integrated nine-chapter monograph.
- `docs/manuscript/01_introduction_and_teleology.md` through `09_tournament_orchestrator_and_inference.md` — chapter-level lineage.
- `docs/empirical_ablation_monograph.md` — 420-match matrix and evaluation caveats.
- `docs/abelian_group_elo_formulation.md` — formal rating proposal.
- `docs/pagerank_and_abelian_graph_invariance.md` — graph analogy.

## Cross-references

- [[pokemon_tcg_docs_corpus_provenance]] — complete document inventory and Git timeline.
- [[pokemon_tcg_stage3_training_failure_postmortem]] — chapter 06 incident expanded.
- [[pokemon_tcg_aug14_ablations_tournaments]] — current denominator-separated results.
- [[pokemon_tcg_empirical_ablation_monograph]] — the empirical monograph's own cohort, deck and numerical interpretation.
- [[pokemon_tcg_data_augmentation_vehicle_draft]] — future vehicle-generation contract derived from the Pilot/Vehicle thesis.
- [[pokemon_tcg_elo_pagerank_invariance]] — rating and graph page.
- [[pokemon_tcg_neural_engine_tokenization]] — chapters 03–06 technical layer.
- [[pokemon_tcg_metanoia_suite]] — adjacent research-protocol layer.
- [[pokemon_tcg_manuscript_chapter_lineage]] — chapter-level provenance and
  current reconciliation.