WikifitaGitHub live67e8de5
pesquisa · kaggle/pokemon_tcg_current_state_reconciliation

Pokémon TCG — Current State Reconciliation and Provenance Ledger

As-of-commit reconciliation of live code, database evidence, handoff blueprints, tests, transcripts and Wikifita claims.

Baixar raw

Pokémon TCG — Current State Reconciliation and Provenance Ledger

Verified snapshot

This page is the current-state anchor for the project. It is not a replacement for the dated timeline or historical experiment pages.

SurfaceVerified valueRole
Project branchdevelopLive implementation checkout
Project HEAD20d7d0dAdds AGENTS.md, the Odyssey Diary and convergence contract
Project commit count144 reachable commits at HEADDirect git rev-list --count HEAD; the older Wiki figure of 147 is not reproducible at this snapshot
Wikifita branchmainPublic knowledge-base checkout
Previous Wikifita incorporation anchor383a0eaCompleted corpus/provenance pass before this coherence pass
Source treeClean before this documentation passNo project-code change authorized or made

The repository and Wikifita are distinct Git histories. A project commit hash proves source evolution; a Wikifita commit hash proves editorial incorporation. Neither replaces the other.

Evidence hierarchy for current claims

For an implementation statement, use this order:

  1. live source code at a named project commit;
  2. executable artifacts, logs, database and dataset probes tied to that commit or snapshot;
  3. project handoff and design documents;
  4. Antigravity reports and transcript artifacts;
  5. older Wikifita pages and user interpretations.

The lower layers preserve intent and reasoning. They cannot override a direct absence in the current source tree.

Current implementation boundary

AreaCurrent verified stateProvenance
MLX trainingscripts/bc/bc_train_mlx.py is FP32-native: it calls model.set_dtype(mx.float32) and rejects non-FP32 model leavesProject HEAD 20d7d0d, source lines around the trainer initialization
PyTorch inferencerl/policy_infer_torch.py converts and validates floating tensors as strict torch.float32Project HEAD 20d7d0d
Runtime action selectionagent/main.py recomputes logits for each multi-select substep, masks picked options and carries memory_out into the next decisionProject HEAD 20d7d0d
Temporal memoryScratch registers provide a memory interface; memory_in and memory_out are carried per tracker state, with TBPTT lane isolationrl/policy_mlx.py, rl/policy_infer_torch.py, agent/main.py
Core architectureEntity/action Transformer, d_model=128, four heads, three-layer default or four-layer session configuration, FFN 512Current source/config; exact run config wins for a checkpoint
Position encodingTyped zones, scalar/meta features and explicit source/target gathers; no tracked RoPE-ND implementationDirect source-tree inventory
MoE / vehicle / ApexNo tracked rl/ropend, rl/moe, policy_moe_*, vehicle_draft.py or Apex runtime pathDirect source-tree inventory
DatabaseSQLite Schema 2.0.0, populated but not clean: foreign-key audit reports 2,946,336 violationsRead-only database probe and pokemon_tcg_aug14_data_etl_database_audit

The FP16 trainer and FP16 PyTorch artifact remain valid historical delivery records. They are not the current source contract after the FP32 correction.

Read-only database revalidation

The native ResultsDB API was used without network synchronization or mutation. The current snapshot returns:

Table or metricValue
matches139,783
match_steps3,065,706
agents / decks1,117 / 948
days / datasets30 / 30
agent_elo_daily / deck_elo_daily28,745 / 10,742
tournaments / matchups128 / 1,621
direct source split138,023 remote / 1,760 local matches
Deck #633 local invariant Elo412.21 over 420 games; MD10 complete
Deck #251 local invariant Elo577.39 over 420 games; MD10 complete

The invariant Elo values are database observations under the current ResultsDB.get_invariant_deck_elo implementation. They do not replace the deck-conditioned tournament win rates, and they should not be interpreted without the source, season and matchup context.

Handoff milestone status

The August 14 handoff commit f988eb0 added PROJECT.md, TEST_INFRA.md, .agents/ survey artifacts, scratch probes and a Tier 1 E2E test file. It did not add the proposed RoPEND, MoE, vehicle-draft or Apex modules.

MilestoneStatus in the project handoffDirect reconciliation
M1 RoPEND/MoEIN_PROGRESS; gate PENDINGBlueprint/explorer material exists; implementation modules are absent
M2 elite dataset/DB parityIN_PROGRESS; survey phase openNo purge/rebuild executed; FK and ZIP/DB deltas remain
M3 PageRank/Abelian monographIN_PROGRESS; gate not passedDocuments exist; the claimed implementation-level isomorphism is not verified here
M4 Wikifita syncPLANNED in the project planWikifita integration has progressed externally, but this does not rewrite the project plan's historical milestone status
M5 E2E/tournament hardeningPLANNEDNo verified 500-match acceptance run with the stated >40% Yan target
E2E test writersInitialized or dispatchedProgress files retain unchecked execution and handoff boxes; generated tests are not proof of passing gates

“In progress” in PROJECT.md or .agents/ means orchestration status. It does not mean that the source modules or acceptance gates exist.

Post-blueprint Git lineage

These commits belong in the project timeline and explain the current handoff state:

CommitChangeEvidence interpretation
f988eb0Added blueprints, milestone scopes, survey artifacts, scratch probes and E2E test scaffoldingPlanning and exploration artifact; not delivery proof
ea9317aIgnored SQLite binary sidecars and synchronized GEMINI.mdRepository hygiene plus memory-state edit
dbf298aRecorded three-tier idempotency sync, bulk Elo decoupling and KaTeX directive changesSmall source/memory refactor; requires code inspection for behavior claims
14ec607Updated channel/rules files, ResultsDB and card-stat build path; added follow-me ruleOperational rule and ETL/Elo maintenance lineage
20d7d0dAdded the Odyssey Diary and multi-agent convergence contract in AGENTS.mdCurrent project instruction/handoff boundary

The earlier documentation pass stopped its provenance table at 12595d4. This ledger closes that gap without rewriting the historical claims those pages recorded.

Antigravity transcript and brain cross-check

The August 14 handoff was not reconstructed from one summary. The main Antigravity session 9189fa2e-93c2-4a04-9bf0-6d090880de27 and its linked brain artifacts were checked against the repository copies and milestone ledgers. The portable transcript location is ~/.gemini/antigravity-cli/brain/<session-id>/.system_generated/logs/, with both compact and full JSONL variants. The repository handoff also identifies the deck analyst 687d75b1-19ff-4999-a56e-c5d40d869f0f, migration engineer 245904ec-c5c7-4dca-a796-e04db0afb510, sentinel da74ca6c-c155-4353-82d1-5c5695a60da1, orchestrator cd851a4f-6875-4819-9f25-1b23dd14cc1b, and M1/M2/M3 milestone sessions.

The cross-check resolves the latest handoff into four evidence classes:

Brain or transcript materialWhat it establishesWhat it cannot establish
fp16_precision_crisis.md and the migration transcriptA reported FP16 export collapse and an FP32 restoration pathThat every historical run used the same artifact or that the report replaces live code
ablation_matrix_results.mdThe 420-game Stage 1-4 FP32 matrix and deck-conditioned peaksThat its rows share one database lifecycle or one later tournament cohort
project_architecture_and_state.mdThe Stage 3 loss-scale narrative, Stage 4/5 chronology and scratch-register interpretationThat scratch registers causally absorbed the shock; this remains an interpretation
master_project_state_and_handoff_plan.md, .agents/ transcripts and milestone ledgersThe intended M1-M5 topology, acceptance criteria and handoff stateCompletion of absent modules, gates, parity or the 500-match threshold

Recent transcript decisions are also preserved in pokemon_tcg_cross_harness_provenance: bulk/idempotent ETL, 28-hour leaderboard TTL, separate Elo recomputation, invariant-Elo preference, docs-only continuation after the quota boundary, and explicit preservation of future architecture. These are decisions and operating constraints; each implementation claim still resolves through the live source and artifact hierarchy above.

Contradictions resolved in Wikifita

The following were genuine stale-reference problems, not legitimate experiment cohorts:

  • FP16 versus FP32: FP16 is retained as the July 25 and August 7 delivery history; strict FP32 is the current source contract.
  • Scratch state: current inference carries memory_out; the old statement that scratch outputs are discarded is no longer current.
  • Multi-select: current agent/main.py performs autoregressive substeps; the old topk(count) statement belongs to the earlier defect record.
  • Commit count: 147 was a previous timeline snapshot; 144 is the direct count at current develop HEAD.
  • Blueprint completion: M1/M2/M3 “in progress” and M4/M5 “planned” remain handoff statuses; absent modules and pending gates control implementation truth.
  • Database cleanliness: the handoff's zero-FK-error and full-parity acceptance target is not current evidence; the direct audit still finds 2,946,336 violations and unresolved archive deltas.
  • Historical architecture sanity versus Stage 3/4 failure: the August 6-7 ten-run suite found no broad encoding, conversion, cache or collapse failure within its own cohort and therefore supported a data- or regime-limited interpretation. The later Stage 3/4 audit found a loss-scale and validation contract defect. These are time-scoped observations, not competing claims about one undifferentiated current run.
  • Prospective sidecar status: the sidecar and lateral planner are historical artifacts removed at a942373; current auxiliary labels and the would-KO contract survive in the primary pipeline. Pages that call the sidecar “current” are stale and must defer to pokemon_tcg_prospective_v2 and pokemon_tcg_research_backlog.
  • Entity normalization status: the L1/L2 ghost-team work is an analysis backed by Antigravity scratch probes, not a tracked production resolver or a persisted identity-mapping table. The prose <2,000 episode heuristic and the saved +60,000 probe window are distinct artifacts; see pokemon_tcg_entity_normalization.

Legitimate differences remain separated: FP32 Stage 1–4 tournament cohorts, deck-conditioned results, teacher sweeps, database resets and historical design snapshots are not collapsed into one ranking.

Cross-references