WikifitaGitHub live67e8de5
pesquisa · kaggle/pokemon_tcg_aug14_architecture_and_handoff_audit

Pokémon TCG AI Battle - 2026-08-14 Architecture and Handoff Audit

Cross-check of the August 14 handoff, live source tree, Antigravity provenance, implementation gates, and preserved future architecture.

Baixar raw

Pokémon TCG AI Battle - 2026-08-14 Architecture and Handoff Audit

Purpose and source order

This page records the August 14 architectural review. It is an additive audit, not a pruning pass. Prospective material such as multidimensional RoPE, MoE routing, vehicle drafting, Apex mode, and later research phases remains documented below even when the live source tree does not implement it.

For implementation claims, the order of authority used here is:

  1. live source code and its executable artifacts;
  2. current repository handoff and project documents;
  3. inspectable training, tournament, database and dataset artifacts;
  4. Antigravity explorer reports and transcripts, labeled as provenance;
  5. older Wikifita pages and historical assumptions.

What the live source tree implements

The current policy implementation is still the entity/action Transformer family in rl/policy_mlx.py and rl/policy_infer_torch.py, trained by scripts/bc/bc_train_mlx.py. The observed contract is:

AreaObserved August 14 state
Model widthd_model=128, four attention heads, monolithic FFN width 512
Recurrent workspaceScratch registers, learned initialization, memory_in and memory_out; the active count varies by configuration and run
SupervisionPolicy/action loss plus aux_ko, aux_prize_delta, aux_terminal and aux_return heads
Temporal trainingTBPTT with per-episode and per-side memory isolation in the trainer
PrecisionThe current PyTorch inference contract validates FP32 tensors; historical FP16 artifacts remain in the repository for lineage and comparison
Position handlingExisting token types, scalar/meta features and explicit option source/target gathers; no explicit RoPE implementation was found
Future modulesNo tracked rl/ropend, rl/moe, rl/policy_moe_torch.py, rl/policy_moe_mlx.py or rl/deck/vehicle_draft.py implementation was found in the current tree

The last row is an absence observation, not a deletion request. The blueprints and tests remain part of the project's future-phase record.

What the August 14 documents specify

The current repository handoff (AGENTS.md, PROJECT.md, TEST_INFRA.md, GEMINI.md, docs/technical_handoff_rfc.md and docs/architecture/) defines a larger execution plan:

  • 4D RoPEND over match step, meta-epoch, urgency clock and inferred Elo;
  • four-expert Top-2 MoE with load-balancing supervision;
  • 60-card vehicle draft context;
  • strict FP32 contracts and a later Apex runtime phase;
  • an elite dataset and relational parity work;
  • mathematical monograph and Wikifita synchronization tracks.

These are current design and execution-plan claims. They are not evidence that the corresponding modules have already landed in the source tree. The August 14 milestone ledgers confirm this distinction: M1 is in implementation with its gate pending, M2 is still in technical survey, and M3 has explorer work but no final gate.

Antigravity provenance cross-check

The Antigravity transcripts were inspected through their local brain records. The repository copies under .agents/ are the durable, reviewable handoffs; the transcript UUIDs identify provenance and do not themselves prove implementation.

Provenance recordRepository handoff or mirrorStatus observed
Main orchestration 9189fa2e-93c2-4a04-9bf0-6d090880de27AGENTS.md, PROJECT.md, TEST_INFRA.md, RFC and generated survey artifactsHandoff and blueprint synthesis
M1 orchestration 9a189410-43b1-4cdc-bc2a-7a942180e59c.agents/sub_orch_m1/Explorers complete; worker and gate pending
M1 RoPEND explorer 607511a6-cff5-4500-bc83-612199d23b85.agents/m1_exp_ropend/handoff.mdMathematical/operator blueprint complete; implementation absent
M1 contract and MoE explorers.agents/m1_exp_contract/handoff.md, .agents/m1_exp_moe/handoff.mdBlueprint handoffs complete; implementation gate pending
M2 orchestration f5143692-4dba-4e8a-aa34-f7465d296f9b.agents/sub_orch_m2/Technical survey still in progress; no purge or elite rebuild executed
M2 DB explorer 85613988-6d11-4bc9-9fac-20d973ad499b.agents/explorer_m2_db/Dispatch initialized, but no completed repository handoff was present
M2 dataset explorer d2204035-8f25-4c8e-8e7b-dde82ad0dccb.agents/explorer_m2_dataset/Progress claims audit work, but no completed handoff file was present
M2 oracle explorer 4f2002e7-7972-4adb-a21f-95806599d6c4.agents/explorer_m2_oracle/handoff.mdOracle report complete; tests reported passing
M3 orchestration 4877bc7d-bfc2-44d3-bc55-1a9dd628ba39.agents/sub_orch_m3/Mathematical explorer reports exist; gate remains in progress

The M2 survey handoff is therefore useful evidence of what Antigravity measured and proposed, but its claims about a completed purge, zero foreign-key errors or a completed elite rebuild are not current implementation facts. The live database audit is recorded separately in pokemon_tcg_aug14_data_etl_database_audit.

Preserved future research boundary

The following remain valid future-phase records and must not be collapsed into the current implementation snapshot:

  • RoPE-ND as a relational phase hypothesis over content carried by scratch and phase coordinates carried by rotations;
  • structured MoE experts and vehicle-level context;
  • causal tests separating recurrent persistence, future decodability, incremental sufficiency and causal mediation;
  • Engram, multi-chain persistence, GRPO/RL alignment, world-model and later topology ideas;
  • the mathematical PageRank-Abelian and J-space research programs.

The correct label for these items is planned, hypothesis, open, or deferred according to the originating document. They are not removed merely because the present model still uses the existing Transformer path.

August 15 corpus expansion

This audit is now the entry point to a deeper documentation graph rather than the endpoint of the integration. pokemon_tcg_docs_corpus_provenance maps all 35 repository documents and their Git history; pokemon_tcg_neural_engine_tokenization records the as-built tensor contract; pokemon_tcg_ropend_moe_blueprint preserves the later RoPEND/MoE, vehicle-draft augmentation and Apex lineage; pokemon_tcg_data_augmentation_vehicle_draft gives that augmentation its own provenance contract; and pokemon_tcg_cross_harness_provenance records how Antigravity reports are used without replacing live source evidence.

The training failure that motivated the user's correction is documented in pokemon_tcg_stage3_training_failure_postmortem. It is part of the architecture boundary: adding prospective heads or future experts without a declared loss denominator, feature lineage and deployment-aligned validation can damage the shared representation while leaving in-pool metrics plausible.

Current reconciliation after the handoff wave

The verified endpoint is project commit 20d7d0d on develop, with 144 reachable commits. The post-blueprint commits f988eb0, ea9317a, dbf298a and 14ec607 add planning, repository hygiene, ETL/idempotency and operational-rule lineage, but they do not establish that RoPEND, MoE, vehicle-draft or Apex runtime modules landed. A direct source and database recheck therefore supersedes any unqualified handoff wording that sounds like an acceptance result. The detailed contradiction ledger is in pokemon_tcg_current_state_reconciliation, which should be read alongside this dated August 14 audit rather than replacing it.

Cross-references