Pokémon TCG — Curriculum Stage Lineage and Cohort Boundaries

Why this page exists

The project uses “Stage 1” through “Stage 5” for several related but non-identical objects: training curricula, checkpoints, source filters, tournament packages and future plans. A stage label is not a cohort key. This page keeps the stage narrative attached to its source artifact, commit, denominator and status so that the 871-game, 420-game, 3,600-game and later teacher results are not silently merged.

Stage lineage

Stage or phase	Source-reported intent	Directly observed boundary	Status in the current reconciliation
Stage 1	Broad behavioral-cloning foundation over the available corpus	FP32 package in tournament 119: 60/420, 14.29%; an earlier 871-game round-robin record reports 28.47%	Historical checkpoint/cohort, not a current acceptance result
Stage 2	Elo-filtered or more selective curriculum, described as Top 600	FP32 package in tournament 120: 64/420, 15.24%; earlier round-robin record reports 28.93%	Historical checkpoint/cohort; filter semantics depend on the run artifact
Stage 3	Top-100 elite curriculum and a continuation from Stage 2	The run labeled one day loaded two dates, split validation inside the filtered pool, and changed the sparse auxiliary-loss reduction; tournament 121 records 58/420, 13.81%	Failure-analysis cohort; loss and validation contract not accepted as a clean baseline
Stage 4	Loss-corrected or stabilized continuation, with a stronger vehicle effect	FP32 package in tournament 122: 72/420, 17.14%; Deck #633 reached 39/140, while Deck #251 reached 18/140	Best aggregate member of the comparable 420-game cohort, still below the handoff target
Stage 5	Top-25 “Apex” polishing before an RL transition	The August 12 architecture report calls it scheduled; no current source path or acceptance artifact proves a completed Stage 5 runtime	Planned/source-reported future phase
Post-BC direction	Offline GRPO, self-play, frozen teacher/world-model experiments and later MoE specialists	The long-horizon vision and August 13 blueprint describe the direction; no corresponding current runtime modules are tracked	Deferred research lineage

The Stage 3 and Stage 4 labels must be read with pokemon_tcg_stage3_training_failure_postmortem. The user-reported catastrophic-forgetting interpretation is preserved as a causal hypothesis; the code and logs establish a loss-scale and validation mismatch, not an isolated gradient-attribution proof.

Cohorts that must remain separate

Cohort	Source and storage	Denominator and result	Why it is separate
Full round-robin	SQLite tournament IDs 98–102 and `project_architecture_and_state.md`	871 games per agent in the report; Stage 1 28.47%, Stage 2 28.93%, Stage 3 epochs 31/32 30.42%/30.19%, teacher 67.16%	Different opponent/deck configuration and lifecycle from the 420-game matrix
Benchmark sweep	SQLite tournament IDs 105–109	20-game benchmark; stages 1–4 from 14.17% to 14.72%, teacher 43.89%	Smaller evaluation and different sweep configuration
Principal FP32 matrix	SQLite tournament IDs 119–122 and `ablation_matrix_results.md`	420 games per stage; 14.29%, 15.24%, 13.81%, 17.14%	Comparable Stage 1–4 matrix used by the empirical monograph
Cross-agent report	`tournament_analysis_2274.md` in the Antigravity brain	3,600-game report with a Stage 4 peak of 27.5% in its own deck/opponent setup	Separate report and matchup construction; not the 420-game matrix
Teacher sweeps	SQLite tournament IDs 123–128	800-game and later rows, including 12.52%, 12.38%, 12.75%, 12.50%, then teacher 51.00% and 64.35%	Different teacher and reset/opponent conditions
Curriculum suite	August 6–7 ten-run suite	`(1d, 3d, 5d) × (1 epoch, 10 epochs) × (top-Elo OFF, ON)`, with per-run tournaments and a 45-pair peer round-robin	Validation and peer/public-agent observables are not the Stage 1–4 FP32 matrix

The database currently preserves tournament summaries but only 1,760 local match rows in the normalized matches table. It therefore does not prove that every earlier tournament's underlying replay rows survived one continuous database lifecycle. The reset and retention uncertainty is part of the result provenance, not a footnote.

Incident inside the lineage

The most important stage transition is not a score increase. It is the change in the training contract around Stage 3:

the run was labeled as a one-day Top-100 curriculum but loaded two dates;
validation was drawn from the same filtered pool rather than a declared deployment-aligned holdout;
commit 0dec40f changed the auxiliary reduction path by removing the valid-row division from the accumulated objective;
the next Stage 4 telemetry exposed training auxiliary values near 1,000 beside validation values near 0.2–1.5;
the reported “dynamic mean” correction changed the scale again, while the strict-FP32 correction addressed a separate numerical failure.

The consequence is methodological: Stage 4's 17.14% cannot be cited as an isolated proof that one loss fix improved the architecture. It is an observed result under a changed objective, checkpoint lineage, precision contract, vehicle mix and tournament cohort.

Provenance chain

Commit or artifact	Transition recorded	Evidence role
`a942373` (2026-08-03)	Removed the separate prospective sidecar, integrated auxiliary heads/meta features and moved the corpus to Parquet	Code change; establishes the post-sidecar baseline
`f90ac1e` (2026-08-11)	Added curriculum-v1 sweep orchestration, model/deck combinations and atomic Elo updates	Code change; orchestration surface
`86aa528` (2026-08-12)	Added Stage 3 tournament and Abelian-rating changes	Code/data change; not proof of a theorem
`0dec40f` (2026-08-12)	Changed the trainer's auxiliary reduction path	Code change; central to the Stage 3 incident
`d9c67a7` (2026-08-12)	Added the integrated monograph and nine manuscript chapters alongside training/tournament edits	Mixed code/document commit; requires claim separation
`9bc5021` (2026-08-12)	Stored Stage 1–3 FP32 archives	Artifact lineage; checkpoint existence only
`459f9d6` and `f697ce9` (2026-08-13)	Introduced the first Magnum Opus/RoPEND/MoE and stochastic-Elo blueprint layer	Future design lineage
`docs/empirical_ablation_monograph.md`	Formalized the 420-game matrix, validation/win-rate distinction and vehicle interpretation	Source monograph; cohort-bound synthesis
`ablation_matrix_results.md`	Reproduced the principal 420-game table in the Antigravity brain	Cross-harness report; cross-checked against SQLite
`project_architecture_and_state.md`	Recorded the 871-game narrative, Stage 3 anomaly and planned Stage 5	Dated handoff; not a current runtime snapshot

The full repository order is maintained in pokemon_tcg_repository_timeline, while the source-to-page mapping is in pokemon_tcg_docs_corpus_provenance.

Reading rule for future experiments

Every new stage report should include:

stage label / checkpoint or commit / source days / filter / objective reduction /
precision / deck set / opponent set / game count / reset boundary / artifact path

If one field is missing, the result remains a dated observation with an explicit uncertainty rather than becoming a new baseline. This contract is the minimum needed before comparing a future vehicle draft, GRPO run or MoE specialist with the historical BC stages.

Cross-references

pokemon_tcg_aug14_ablations_tournaments — stored tournament ledger and denominator-separated results.
pokemon_tcg_empirical_ablation_monograph — analytical reading of the principal matrix.
pokemon_tcg_bc_curriculum_ablation — ten-run curriculum suite.
pokemon_tcg_stage3_training_failure_postmortem — loss-scale and validation incident.
pokemon_tcg_cross_harness_provenance — Antigravity report and transcript lineage.
pokemon_tcg_ropend_moe_blueprint — future architecture that depends on repaired stage contracts.
pokemon_tcg_long_horizon_vision — post-BC GRPO, self-play and MoE direction.
pokemon_tcg_current_state_reconciliation — current code/database boundary.