Pokémon TCG - Fita Labs August 14 Source Set

Purpose

The August 14 Fita Labs source set is a three-document analytical package supplied alongside the repository handoff. The documents are related, but they are not interchangeable:

1. the Codex handoff is an operational and strategic orientation;
2. the analytical thesis is a conceptual and mathematical research program;
3. the TBPTT/J-space dossier is an evidence-bound reconstruction of the recurrent-workspace observation and its unresolved causal confounds.

This page keeps those roles visible in Wikifita. It does not turn a thesis claim into a runtime fact, and it does not discard later phases because they are not implemented at the current repository endpoint.

Source inventory

Source	Version and size	Primary role	Editorial status
Fita_Labs_Handoff_Pokemon_Codex_v1_2.pdf	v1.2, 14 pages, dated 2026-08-14	Operational handoff, competition strategy, architecture map, ETL direction, acceptance questions and deferred work	Handoff source; current code, logs and database outrank its unverified completion language
Fita_Labs_Analytical_Thesis_v0_5.pdf	v0.5, 22 pages, dated 2026-08-14	Deterministic chained memory, fibred trajectories, recurrent workspace, provenance, adaptive compute and J-space-like research	Conceptual thesis; falsifiable research program, not mechanistic proof
Fita_Labs_TBPTT_JSpace_Evidence_Dossier_v1_1.pdf	v1.1, 12 pages, dated 2026-08-14	Evidence ladder for TBPTT, persistent scratch, future-predictive auxiliary heads, normalization confound and RoPE-ND follow-up	Analytical dossier; current logs and code establish the observed boundary, while causal interpretation remains open

The PDFs are user-supplied source artifacts rather than Git-tracked project files. Their filenames, versions and dates are retained here so the claims can be traced back to the supplied source set without treating a derivative wiki summary as the original.

Shared claim boundary

The three sources converge on a narrower and more useful conclusion than the strongest language in any one handoff paragraph:

pre-TBPTT baseline
    -> TBPTT plus persistent scratch
    -> auxiliary heads learn future-relevant signals
    -> policy objective and validation become confounded
    -> RoPE-ND becomes a separate relational hypothesis
    -> causal probes and ETL/validation repair remain open

The current repository independently supports the following implementation boundary: recurrent scratch and TBPTT exist; auxiliary future/terminal/return heads exist; the current inference contract is strict FP32; no explicit RoPE, RoPE-ND, MoE, vehicle-draft or Apex runtime was found in the tracked source tree. The absence of a current module does not delete its documented future phase.

Handoff reading

The Codex handoff frames the immediate mission as producing a competitive submission while preserving enough provenance to make later interventions auditable. It explicitly warns that the source is iterative and that current code, configuration, logs and checkpoints outrank stale documentation.

Its architecture map separates the current recurrent behavioral-cloning system from later interventions. In that map, ETL throughput and data quality may be the bottleneck; future work includes isolated RoPE-ND, recurrent workspace probes, better validation, data/ETL expansion, GRPO or self-play and later MoE or world-model directions. Those items remain in pokemon_tcg_long_horizon_vision, pokemon_tcg_ropend_moe_blueprint and pokemon_tcg_data_augmentation_vehicle_draft.

The handoff's acceptance questions are useful as a checklist, not as evidence that the gates passed. They require explicit scratch persistence, exact TBPTT detach semantics, loss denominators, future-head metrics, disjoint data boundaries, replay/ETL inventory, tournament protocol and current positional encoding. The August 14 repository audit records the corresponding gates as pending where the live source or artifacts do not establish completion.

Thesis reading

The analytical thesis distinguishes several coordinates that should not be collapsed:

Coordinate	Working meaning	Pokémon TCG evidence boundary
J_t	Compact recurrent scratch or working-state coordinates	Future-relevant information is reported from auxiliary heads; causal policy mediation is open
H_t	Provenance or lineage heads and recoverable history	The project has replay, database and artifact provenance, but not a proved internal provenance manifold
Phase or relational coordinates	Explicit relations among states, events or positions	RoPE-ND is a later experiment; current token types and scalar/meta features are not RoPE-ND
Fibred trajectory	A state trajectory with admissibility, provenance and temporal structure	A research abstraction for organizing replays and interventions, not a claim about a completed runtime topology

The thesis deliberately separates semantic reuse from provenance, gradient horizon from working-memory horizon, and working-memory horizon from provenance horizon. That separation is compatible with the current project architecture, but it does not prove that a hidden model state is identical to the proposed mathematical object.

The thesis also treats identity, agency and J-space as system-level or operational hypotheses. Wikifita retains those ideas as research vocabulary and falsifiable questions, not as evidence of sentience, selfhood or a particular internal latent manifold. The general J-space framing remains in j_space_inference.

TBPTT and J-space evidence dossier

The dossier is the most direct source for the recurrent-workspace correction. Its evidence ladder is:

1. a pre-TBPTT model existed;
2. TBPTT and persistent scratch were introduced;
3. the policy result deteriorated under a confounded training regime;
4. auxiliary heads nevertheless learned future-relevant quantities from the recurrent state;
5. this observation predates RoPE-ND;
6. the result is compatible with a J-space-like functional workspace, while causal policy contribution and formal equivalence remain open.

The dossier's proposed controls are therefore essential: verify persistence across the actual TBPTT boundary, compare future probes with and without recurrent scratch, reset or patch scratch state, measure policy effects, compare normalized auxiliary objectives and isolate a minimal RoPE-ND phase intervention. These controls complement the Stage 3/4 postmortem rather than replacing it.

Stage 3 incident and validation contract

The source set gives a richer account of the failure already recorded in pokemon_tcg_stage3_training_failure_postmortem:

• the run labeled as a one-day Top-100 curriculum loaded two dates;
• validation was drawn from the same filtered pool rather than a declared deployment-aligned holdout;
• sparse auxiliary reduction and optimizer normalization did not share one declared denominator;
• training and validation auxiliary values became incomparable;
• the shared trunk could receive disproportionate auxiliary pressure while auxiliary heads continued to emit useful signals;
• catastrophic forgetting is a credible mechanism proposed by the user and supported by the pattern, but not an isolated gradient-attribution proof.

The “pipeline mixing” language in the source set is recorded as a contract failure and audit requirement. It is not expanded into a claim of a specific row-level leak until the dataset builder, split construction and feature lineage are traced for the exact run.

RoPE-ND and later phases

The source set changes the motivation for RoPE-ND. It is not the missing ingredient required for recurrence to carry temporal information, because the reported future-predictive auxiliary behavior occurred before RoPE-ND was present. Its remaining research question is narrower: whether explicit phase or relational coordinates improve transport, addressability or intervention over content already carried by recurrent scratch.

The following future records remain intentionally preserved:

• RoPE-ND or multidimensional RoPE as an isolated relational experiment;
• future data augmentation and vehicle-draft generation;
• GRPO, self-play and a frozen-teacher/world-model phase;
• MoE specialists, Apex mode and later routing experiments;
• Engram, multi-chain provenance and the broader J-space mathematical program.

Their status remains planned, hypothesized, deferred or open until live code, artifacts and tests establish otherwise. See pokemon_tcg_current_state_reconciliation and pokemon_tcg_curriculum_stage_lineage for the current boundary.

Source-to-Wikifita map

Source-set material	Canonical Wikifita location
Operational handoff and acceptance gates	pokemon_tcg_aug14_architecture_and_handoff_audit, pokemon_tcg_handoff_acceptance_reconciliation
Repository docs and Git provenance	pokemon_tcg_docs_corpus_provenance, pokemon_tcg_repository_timeline
Stage chronology and cohort separation	pokemon_tcg_curriculum_stage_lineage, pokemon_tcg_aug14_ablations_tournaments
Loss-scale and validation incident	pokemon_tcg_stage3_training_failure_postmortem
Current recurrent/token contract	pokemon_tcg_agent_architecture, pokemon_tcg_neural_engine_tokenization, pokemon_tcg_training_pipeline
Future RoPE-ND, MoE and vehicle architecture	pokemon_tcg_would_ko_prospective_search, pokemon_tcg_ropend_moe_blueprint, pokemon_tcg_data_augmentation_vehicle_draft, pokemon_tcg_long_horizon_vision
J-space and recurrent-workspace research	j_space_inference, pokemon_tcg_metanoia_suite, pokemon_tcg_manuscript_and_empirical_synthesis
Cross-harness and Antigravity provenance	pokemon_tcg_cross_harness_provenance, pokemon_tcg_orchestration_artifact_ledger

Editorial rule

When a future agent reads one of these PDFs, it should add evidence to the appropriate page and preserve the source role. Do not flatten the handoff, thesis, dossier, code, logs, tournament cohorts and database snapshot into a single status label. Keep claims classified as observed, code-validated, source-reported, inferred, hypothesized or unresolved.