WikifitaGitHub live67e8de5
pesquisa · kaggle/pokemon_tcg_cross_harness_provenance

Pokémon TCG — Cross-Harness and Antigravity Provenance

Sanitized ledger for Antigravity transcripts, report artifacts, tokenizer/channel mappings and the evidentiary role of cross-harness handoffs.

Baixar raw

Pokémon TCG — Cross-Harness and Antigravity Provenance

Why this ledger exists

The August 14 handoff was assembled across repository work, local artifacts and Antigravity sessions. Those surfaces are provenance sources, not interchangeable authorities. A transcript can explain why a hypothesis was proposed; the live repository and run artifacts determine whether it is implemented or reproduced.

The repository's docs/cross_harness_and_tokenizer_adaptation.md is the canonical project-relative ledger. This page makes its role visible in Wikifita without exposing machine-specific absolute paths.

Primary Antigravity lineage

The main handoff is associated with the Antigravity research identity:

9189fa2e-93c2-4a04-9bf0-6d090880de27

The local transcript family is stored under the Antigravity brain/transcript tree, with report artifacts addressed by their session or artifact identifier. The portable form is:

~/.gemini/antigravity-cli/brain/<session-or-artifact-id>/

The exact filesystem ledger remains in the repository handoff so a future agent can resolve the report without treating a copied Wikifita page as the source of truth.

Key report artifacts

ArtifactWhat it contributesEvidence status
project_architecture_and_state.mdStage chronology, scratch-register interpretation, Stage 1–3 round robin and Stage 4 transitionSource-reported handoff; cross-check required
ablation_matrix_results.mdFP32 Stage 1–4 420-match matrix and deck-conditioned peaksMatches the current tournament ledger for the principal matrix
fp16_precision_crisis.mdMLX versus PyTorch FP16 collapse and FP32 restoration narrativeEmpirical artifact; runtime contract must be checked separately
tournament_analysis_2274.mdA separate tournament analysis with catastrophic-regression framingSeparate cohort; do not merge into the 420-match matrix
.agents survey artifactsFilesystem, dataset, run and handoff inventorySurvey evidence; unresolved count deltas preserved

The Antigravity report says the Stage 3 dynamic-mean anomaly produced a loss near 81.0, concentrated sparse auxiliary gradients and a subsequent Stage 4 pure-sum correction. The current trainer history independently confirms a reduction change and the Stage 4 training/validation scale discontinuity. The report's stronger statement that scratch registers absorbed the shock remains an interpretation.

Additional main-brain artifacts

The main session also produced artifacts that are easy to mistake for implementation or acceptance evidence:

ArtifactWhat it recordsCorrect status
master_project_state_and_handoff_plan.mdLocked Meta roadmap for August 16–31, component inventory, invariant-Elo equations and verification checklistFrozen handoff blueprint; it contains source-reported “100% parity” and “PageRank isomorphism” language that must be reconciled against the live audit
prompt_draft.mdLaunched teamwork requirements R1–R4 and the strict-FP32, 500-match, database-parity and Wikifita acceptance criteriaNormative request/acceptance contract; unchecked criteria remain open
schema_mermaid.mdEntity-relationship view of the legacy/current SQLite table families and foreign-key topologyStructural artifact; current counts and integrity status come from ResultsDB and the read-only audit
kaggle_timezone_derivation.mdEpisode-ID boundary interpolation and a UTC cutoff estimate near 00:20:42Quantitative inference; not an official platform contract
.metadata.json sidecarsUser-facing summaries and timestamps for the brain reportsArtifact metadata only; not independent evidence

The full local brain tree contains scratch probes and uploaded media as well. The project documents index the report and transcript roots, while this page records only artifacts that contribute a recoverable project claim. A scratch probe must be cited by path, output and commit or it remains an exploratory trace rather than a fact.

Transcript decision checkpoints

The transcript JSONL and report artifacts were cross-checked again against the repository history. The table records decision provenance, not private reasoning. A decision becomes a project-state claim only when a matching code, artifact or Git change exists.

DateTranscript lineageDecision or operating constraintRepository corroboration
2026-08-13Main session 9189fa2e-93c2-4a04-9bf0-6d090880de27Scan replay/archive structure in bulk; keep ETL idempotent; initialize leaderboard access once; honor a 28-hour CSV TTL; recompute Elo in a separate pass and treat invariant Elo as the operational ratingdbf298a, 14ec607, GEMINI.md, rl/results_db.py and the current ETL/rating pages
2026-08-14Main session 9189fa2e-93c2-4a04-9bf0-6d090880de27Preserve RoPE-ND, MoE, vehicle-draft, Apex and later research even when absent from the live tree; document the handoff rather than silently implementing or deleting itf988eb0, 20d7d0d, PROJECT.md, TEST_INFRA.md, AGENTS.md and pokemon_tcg_current_state_reconciliation
2026-08-14Main session 9189fa2e-93c2-4a04-9bf0-6d090880de27Stop the teamwork-preview work at the compute/quota boundary and leave the frozen handoff intact for a later harnessa71318e, 20d7d0d and the handoff artifacts under .agents/
2026-08-14Main session 9189fa2e-93c2-4a04-9bf0-6d090880de27Keep the immediate continuation docs-only: record current state, transcript paths and provenance without cleaning or repairing project artifacts20d7d0d plus the unmodified project source/database snapshot
2026-08-14da74ca6c-c155-4353-82d1-5c5695a60da1, cd851a4f-6875-4819-9f25-1b23dd14cc1bRun the M1/M2/M3/E2E milestone topology with explicit gates and acceptance criteria.agents/orchestrator_1/, .agents/sub_orch_m1/, .agents/sub_orch_m2/, .agents/sub_orch_m3/ and .agents/orch_e2e/; gate files remain pending or in progress

The relevant local transcript roots are intentionally represented in portable form: ~/.gemini/antigravity-cli/brain/<id>/.system_generated/logs/transcript.jsonl and the corresponding transcript_full.jsonl. The repository handoff maps the main session plus the deck, migration, sentinel, orchestrator, M1, M2 and M3 subagent IDs. These IDs identify provenance and do not establish that a subagent's requested implementation or acceptance gate completed.

Quota boundary and harness transition

The final user checkpoint in the main transcript records that the teamwork preview stopped during active work when the available compute quota was exhausted. The requested continuation was documentation-only: preserve the source tree, database and future architecture, explain why work stopped, and leave enough filesystem provenance for a later harness to resume. The user also explicitly noted that the tokenizer-aware channel protocol and teamwork preview are Antigravity-specific behaviors, not guarantees transferable to another model or harness.

Repository corroboration is a71318e, 12595d4 and 20d7d0d, plus the frozen handoff artifacts. This is a stop-state and a harness-boundary claim; it does not mean that the Locked Meta roadmap, M1–M3 gates or the 500-match acceptance run completed before the stop.

Snapshot separation inside the brain artifacts

The artifacts must remain separate cohorts:

ArtifactSnapshot or cohortCorrect Wikifita treatment
ablation_matrix_results.mdStage 1-4 FP32 matrix, 420 games per stagePrincipal 420-game tournament evidence; cross-check with the stored ledger
project_architecture_and_state.mdAugust 12 curriculum narrative, including Stage 3 shock and Stage 4/5 progressionSource-reported chronology and interpretation; not a current implementation snapshot
tournament_analysis_2274.mdSeparate 3,600-game cross-agent analysisPreserve as a separate cohort; do not merge its rates into the 420-game matrix
fp16_precision_crisis.mdFP16 export collapse and FP32 restoration narrativeHistorical empirical artifact; current FP32 contract is independently code-validated
master_project_state_and_handoff_plan.mdAugust 14 handoff blueprint and acceptance planDesign/handoff status; acceptance bullets remain pending unless direct evidence exists

This separation explains why apparently incompatible win rates, stage labels or claims about “current” architecture can coexist without being edited into one synthetic result. The current reconciliation page is the join point, not a replacement for any dated artifact.

Channel and tokenizer adaptation

The adaptation guide maps conceptual channels and tool behavior across Gemini, Claude, DeepSeek and OpenAI harnesses. Its practical value is preserving the same research object when a handoff changes runtime:

  • channel names and control markers may differ;
  • tokenizer boundaries and special-token assumptions may differ;
  • tool calls can produce different artifact paths or serialization;
  • a model-family adherence observation is not automatically a project-state fact;
  • transcript context can be redacted or summarized without changing the repository's source hierarchy.

The project should therefore carry a provenance tuple with each cross-harness claim:

source harness
session/artifact id
repository commit
artifact path or report name
claim type
verification state

What the transcript search did and did not establish

The available Antigravity reports contain the dynamic-mean/loss-scale explanation and the Stage 1–4 result tables. Searching the available report artifacts did not find a direct phrase proving “validation mixing” or a row-level leakage mechanism. The user's account that Stage 3 mixed signals and lacked correct validation is therefore recorded as an important project interpretation, supported by the in-pool split and scale mismatch, but not falsely attributed to a raw transcript that does not contain it.

This distinction matters because a handoff can compress a debugging history into a causal sentence. Wikifita preserves both the sentence and the verification status.

Handoff reading order

For future analysis, read in this order:

  1. live repository code and current Git commit;
  2. docs/technical_handoff_rfc.md for the map and declared chronology;
  3. docs/cross_harness_and_tokenizer_adaptation.md for the physical artifact ledger;
  4. the named Antigravity report artifacts;
  5. current database, Parquet, logs and tournament records;
  6. Wikifita synthesis pages, including this one.

The order prevents a derivative summary from silently becoming the source for reconstructing its own claims.

Primary sources

  • docs/cross_harness_and_tokenizer_adaptation.md — channel/tokenizer mapping and transcript filesystem ledger.
  • docs/technical_handoff_rfc.md — master handoff and report index.
  • Antigravity report artifacts listed above — source-reported chronology and analysis.
  • docs/metanoia/02_rule_provenance_and_epistemic_evolution.md — rule and directive lineage across harnesses.
  • docs/metanoia/03_model_adherence_and_failure_mode_analysis.md — model-family adherence claims and limits.

Cross-references