Pokémon TCG — Handoff Claims and Acceptance Reconciliation

Why this page exists

The August 14 handoff contains several kinds of text at once:

normative instructions for incoming agents;
a future execution plan and interface specification;
a test-inventory document;
source-reported research state and hypotheses;
acceptance criteria that still require execution.

These roles are not interchangeable. A line in PROJECT.md can define the intended path rl/ropend/; it cannot prove that the path exists. A checkmark in TEST_INFRA.md can describe planned coverage; it cannot replace a test run tied to the current tree. A transcript can explain why the acceptance target was chosen; it cannot satisfy the target by itself.

Source roles

Source	Role in the project	Safe Wikifita interpretation
`AGENTS.md`	Incoming-agent instructions, Odyssey Diary and handoff narrative	Normative instructions plus historical claims; current source evidence still controls implementation status
`GEMINI.md`	Research memory, operating rules and dated project state	Project-memory record; reported proofs and counts require reconciliation
`PROJECT.md`	Magnum Opus feature inventory, milestones, interface contracts and target paths	Future execution plan with declared statuses; not a delivery manifest
`TEST_INFRA.md`	Requirement-derived test architecture and coverage inventory	Declared test design; execution and implementation coverage must be checked separately
`.agents/ORIGINAL_REQUEST.md`	User mission and strict acceptance criteria	Acceptance contract; unchecked boxes remain unchecked until directly verified
`docs/technical_handoff_rfc.md`	Master index for transcripts, runs, database, architecture and manuscripts	Handoff map and source locator; claims retain their source-reported status
`docs/metanoia/01..06`	Research protocol, governance and philosophy	Conceptual and procedural layer, not proof of model cognition or runtime delivery

Milestone status at the verified source snapshot

The source repository is develop at 20d7d0d with 144 reachable commits. The current implementation check and the handoff status are different fields:

Milestone	Declared handoff status	Current evidence	Reconciled status
M1 RoPEND/MoE	`IN_PROGRESS`	Blueprint and explorer handoffs exist; no tracked `rl/ropend`, `rl/moe`, MoE policy or vehicle-draft module	In progress, implementation gate pending
M2 elite dataset/DB parity	`IN_PROGRESS`	Broad Parquet/SQLite corpus and oracle reports exist; direct FK audit still reports 2,946,336 violations; no purge/rebuild executed	In progress, parity gate not met
M3 PageRank-Abelian monograph	`IN_PROGRESS`	Monograph exists and is indexed; code and data support the rating surface, while the stronger spectral isomorphism remains a research claim	In progress, theorem/implementation gate pending
M4 Wikifita sync	`PLANNED` in `PROJECT.md`	The canonical wiki integration is being performed iteratively in this goal and current audit passes	Project milestone label remains planned; current editorial work is observable progress, not a final project acceptance claim
M5 E2E tournament/hardening	`PLANNED`	No verified 500-match acceptance run or `>40%` Yan target in the current evidence set	Planned, not met

The word “in progress” is retained because it is the handoff's state. It is not rewritten as “complete” merely because a document or explorer artifact exists.

Acceptance criteria versus evidence

Criterion from `.agents/ORIGINAL_REQUEST.md`	Current evidence	Result
Strict FP32 checksum and static-feature SHA256	Current trainer and converter enforce strict FP32; a 100% checksum acceptance result is not independently recorded here	Partial contract, acceptance unverified
500 matches against `first_sub_kaggle_2707` without hangs, NaN or leaks	Tournament ledger has multiple cohorts, but no direct current acceptance run tied to this criterion	Not verified
More than 40% overall against `first_sub` on Yan #633	Stage 4 #633 slice is 39/140 = 27.9% in the reconciled cohort	Not met in the observed cohort
SQLite parity and zero FK errors	139,783 current matches, with 2,946,336 FK violations in the read-only audit	Fails current acceptance condition
Wikifita double audit with zero broken links/orphans	`uv run scripts/wikifita_audit.py --fix` and the no-`--fix` pass both pass after this editorial commit	Met for the current Wikifita tree
RFC and Metanoia suite fully indexed and synchronized	All six Metanoia pages and the RFC lineage are indexed in the project/wiki map	Documentation index condition met; this does not complete M1–M3

The observed failure of the database criterion is not a reason to delete the target or repair the database during a documentation pass. It is a status fact that remains linked to the ETL audit and the no-destructive-action boundary.

Feature inventory versus live paths

PROJECT.md lists 16 features. A direct source-tree inventory at the verified endpoint finds the following boundary:

Feature family	Declared path or behavior	Live source observation
4D RoPEND	`rl/ropend/ropend_torch.py`, `rl/ropend/ropend_mlx.py`	Paths absent; the E2E file contains reference implementations instead
MoE routing and experts	`rl/moe/`, `policy_moe_*`	Paths absent
Vehicle draft	`rl/deck/vehicle_draft.py`	Path absent; the vehicle remains a blueprint and data contract
Apex runtime	Date-gated runtime token	No verified Apex runtime path
FP32 contract	Trainer and PyTorch converter	Present and independently source-validated
Elite dataset and auxiliary targets	Parquet/SQLite/oracle pipeline	Broad current corpus and target fields exist; elite recompile and parity acceptance remain open
PageRank/Abelian work	`docs/pagerank_and_abelian_graph_invariance.md`	Document exists; stronger isomorphism claim remains unvalidated as an implementation result
Wikifita audit	`uv run scripts/wikifita_audit.py`	Current wiki audit passes twice

The absence rows are not deletion instructions. The later architecture documents and their future paths remain part of the project's forward lineage.

Test-inventory trap

The current tests/e2e/ directory contains __init__.py and one test_tier1_features.py. That file describes all 16 feature names, but it also contains canonical reference implementations and fallback behavior. For example, its RoPEND test helper attempts to import rl.ropend.ropend_torch and otherwise returns ref_ropend_4d_torch from the test file itself.

This has a precise implication:

declared test coverage != executed implementation coverage

The test inventory can be preserved as a useful contract and scaffold. It cannot be cited as proof that the absent modules exist, that Tiers 2–5 were executed, or that all 184 planned tests passed. The current reconciliation therefore treats E2E completion as pending unless a runnable report with implementation imports, command, commit and result is present.

Claims that require semantic downgrading

The handoff contains strong language such as “proved the isomorphism”, “100% physical parity” and “zero foreign key errors”. The current evidence supports weaker, explicit formulations:

the repository contains a PageRank/Abelian monograph and a formal rating proposal;
the current SQLite snapshot is populated and provides the data surface for the comparison;
the direct FK audit observes 2,946,336 violations;
the 139,783-match count is current for the inspected database snapshot, while archive and database counts have unresolved deltas;
the mathematical equivalence is a research claim or analogy until its assumptions and implementation path are independently validated.

This is a reconciliation of evidence classes, not a correction of the source documents. The source docs remain intact as dated handoff material.

Antigravity and transcript provenance

The handoff's physical ledger points to the portable Antigravity brain form:

~/.gemini/antigravity-cli/brain/<session-id>/.system_generated/logs/transcript.jsonl
~/.gemini/antigravity-cli/brain/<session-id>/.system_generated/logs/transcript_full.jsonl

The main session is 9189fa2e-93c2-4a04-9bf0-6d090880de27. The M1, M2 and M3 orchestration IDs and report roles are preserved in pokemon_tcg_cross_harness_provenance. These records establish decision provenance and reported status. They do not satisfy an acceptance criterion without a matching code path, artifact, test result or database observation.

Reading order for incoming work

Read the current project instructions and verify the designated checkout.
Read PROJECT.md, TEST_INFRA.md and .agents/ORIGINAL_REQUEST.md as plan and acceptance sources.
Read docs/technical_handoff_rfc.md and the named architecture/data/manuscript documents.
Resolve each implementation claim against the live source tree and current artifacts.
Use Antigravity transcript paths and brain reports to recover decision context.
Update Wikifita with claim class, commit hash, cohort, uncertainty and cross-links.

Primary sources

PROJECT.md — feature inventory, milestones and interface contracts.
TEST_INFRA.md — declared E2E test coverage and tiers.
.agents/ORIGINAL_REQUEST.md — mission and acceptance criteria.
AGENTS.md and GEMINI.md — handoff instructions and dated project memory.
docs/technical_handoff_rfc.md — master handoff index.
f988eb0, a71318e, 12595d4, 20d7d0d — blueprint, quota, transcript-ledger and Odyssey handoff commits.

Cross-references

pokemon_tcg_current_state_reconciliation — current code, database, milestone and contradiction ledger.
pokemon_tcg_aug14_architecture_and_handoff_audit — August 14 architecture and Antigravity audit.
pokemon_tcg_cross_harness_provenance — transcript and brain artifact provenance.
pokemon_tcg_orchestration_artifact_ledger — .agents topology, explorer handoffs and pending gates.
pokemon_tcg_aug14_data_etl_database_audit — direct database, archive and FK evidence.
pokemon_tcg_empirical_ablation_monograph — cohort-separated empirical claims.
pokemon_tcg_curriculum_stage_lineage — stage labels, commit lineage and denominator-safe cohort map.
pokemon_tcg_data_augmentation_vehicle_draft — future vehicle and synthetic-data boundary.
pokemon_tcg_metanoia_rule_provenance — rule and epistemic provenance layer.