# Wikifita — Log

Chronological log of wiki operations.

## [2026-09-04] fix | CAMDOM Mikael/Maso Heck identity reconciliation

Corrected the CAMDOM people graph so Mika, Mikael, and Mikaelzinho consistently identify Mikael, Alefita's husband and technical collaborator. Maso Heck is retained only as the Innocean Berlin designer who supplied PSD materials; code, architecture, testing, translations, documentation, and support are attributed to Mikael according to Alefita's clarification and the CAMDOM repository history.

### Pages Updated

- [[camdom-people]] — separated Maso Heck from Mikael and corrected roles, relationships, and credits.
- [[mikael-partner]] — established the identity boundary and CAMDOM contributions.
- [[camdom-development-timeline]] — corrected team and design-phase references.
- [[camdom-scope-and-delivery]] — corrected German-translation attribution.
- [[camdom-debugging-stories]] — corrected the debugging-team reference.
- [[camdom-code-quality]] — existing Mikael contributor count retained as repository evidence.

### Sources

- Alefita's explicit identity and contribution clarification on 2026-09-04.
- Local CAMDOM clone history for `mika-marcondes`, recording 46 Mikael contributions across the repository's Mikael identities.

## [2026-08-15] fix | Pokémon TCG - Harness Session Provenance satellite locators

Rechecked the two Antigravity satellite locators carried forward as unresolved
in the initial `b26c5bd` snapshot. Commit `a9f8cd5` corrected the provenance
page after both brain roots and transcript locators were found locally: the
current classification is 27/27 satellites present. The 19 Claude histories
remain classified from their JSONL `entrypoint` fields as 4 1P and 15 3P.
No Pokémon TCG project file was modified.

### Correction

- [[pokemon_tcg_harness_session_provenance]] — corrected the two satellite
  locator statuses and the aggregate 27/27 statement.
- [[pokemon_tcg_decision_trail]] — clarified the historical policy-path cleanup
  directive against the files still tracked in project commit `20d7d0d`.

### Provenance order

The Git order is `b26c5bd` (initial ledger) followed by `a9f8cd5`
(satellite-locator correction).

## [2026-08-15] create | Pokémon TCG - Harness Session Provenance

Created a portable session provenance ledger for the August 12-14 Antigravity
handoff and the 19 Claude project histories. The initial snapshot records the
transcript-to-artifact/report-to-code/database/Git evidence progression, the
principal session metrics, 27 satellite classifications, the 1P/3P
`entrypoint` split and two locators that were unresolved at creation time,
without reproducing private transcript content. The later `a9f8cd5` correction
resolved those two locators in the current page.

### Pages Created

- [[pokemon_tcg_harness_session_provenance]] - session IDs, scopes, portable
  transcript locators, evidence roles and acceptance boundaries.

### Pages Updated

- `index.md` - Kaggle domain entry and tag coverage.
- `memorias/MEMORY.md` - project-level recoverability pointer.
- `memorias/projetos/pokemon_tcg/README.md` - canonical reading map entry.

### Sources

- Antigravity main and satellite brain transcript metadata and artifact roots.
- Claude project JSONL histories and their `entrypoint` fields.
- Project `.agents` briefings and the existing provenance, orchestration,
  audit, ablation, corpus and decision-trail pages.

### Evidence Boundary

The handoff remains a provenance record. M1-M3, E2E, database parity and the
500-match acceptance gate remain pending; no Pokémon TCG project file was
modified.

## [2026-08-15] deepen + reconcile | Pokémon TCG - implementation, schema, arena, entity and manuscript lineage

Continued from Wikifita anchor `e6c8880` after a fresh verification of project
`develop` at `20d7d0d`, the 35-file `docs/` corpus, the August 14 Antigravity
brain/transcript snapshot through step `4939`, and the current neighboring wiki
pages. This pass expands the graph rather than flattening the repository into
one current-architecture summary.

### Pages created

- [[pokemon_tcg_implementation_specification]] — target service boundaries,
  storage invariants, deterministic workflows and acceptance gates reconciled
  against the partial live implementation.
- [[pokemon_tcg_schema_evolution]] — current physical SQLite debt versus the
  relational v2 identity, replay, experiment and rating design.
- [[pokemon_tcg_arena_future_architecture]] — synchronous local arena and
  future application/worker seam, with deliberate non-goals preserved.
- [[pokemon_tcg_entity_normalization]] — L1/L2 ghost-team analysis, source
  statistics, saved-probe discrepancy and residual identity uncertainty.
- [[pokemon_tcg_manuscript_chapter_lineage]] — nine-chapter Git origin and
  chapter-level routing from narrative claims to current evidence.

### Pages updated

- [[pokemon_tcg_ai_battle]], [[pokemon_tcg_current_state_reconciliation]],
  [[pokemon_tcg_docs_corpus_provenance]], [[pokemon_tcg_repository_timeline]]
  and the project indexes — added the new nodes and their provenance joins.
- [[pokemon_tcg_local_platform_design]], [[pokemon_tcg_local_platform_status]],
  [[pokemon_tcg_physical_schema]], [[pokemon_tcg_replay_data_model]],
  [[pokemon_tcg_dashboard_and_tournament]],
  [[pokemon_tcg_experiments_and_anamnesis]] — connected the implementation,
  schema and future-arena contracts.
- [[pokemon_tcg_data_oracle_etl]],
  [[pokemon_tcg_sampling_timezone_identity]],
  [[pokemon_tcg_cross_harness_provenance]],
  [[pokemon_tcg_data_pipeline]], [[pokemon_tcg_action_coverage]],
  [[pokemon_tcg_ropend_moe_blueprint]],
  [[pokemon_tcg_data_augmentation_vehicle_draft]],
  [[pokemon_tcg_manuscript_and_empirical_synthesis]] and the MLX migration
  page — qualified historical sidecar/FP16 language and added cross-links.

### Source and provenance anchors

- `05c4435` — implementation specification, schema evolution and future arena
  documents.
- `d9c67a7` — integrated monograph and nine manuscript chapters.
- `cec2eab` — empirical ablation specification companion.
- `eb83bef` — normalization heuristics and ETL analysis.
- `20d7d0d` — live project endpoint used for current-state checks.
- Antigravity main brain `9189fa2e-93c2-4a04-9bf0-6d090880de27`, including
  `hierarchical_unifier.py`, `timeline_chain_v2.py` and the handoff reports;
  these remain analytical artifacts, not a production resolver.

### Evidence boundary

The normalization page records the discrepancy between the prose `<2,000`
episode heuristic and the saved `+60,000` exploratory probe window. The new
implementation pages preserve future phases and do not authorize migration,
database rebuild, source-code change, training or tournament execution. No
Pokémon TCG code, model or database was modified.

## [2026-08-15] reconcile | Pokémon TCG - historical cohort claims and sidecar boundary

Reopened the project coherence pass from the live roots after the previous
incorporation commit `383a0ea`. The project remains at `20d7d0d` on `develop`,
and the main Antigravity brain/transcript snapshot remains the August 14
session through step `4939`. Cross-reading exposed stale wording that was
semantically misleading even where the underlying historical observations were
valid.

### Corrections

- Scoped “data-limited, not architecture-limited” to the August 6-7 ten-run
  suite. The later Stage 3/4 loss-scale and validation-contract incident is
  now explicitly linked wherever the older sanity result appears.
- Reclassified Prospective V2 sidecar and lateral-planner references as
  historical after `a942373`; current auxiliary heads and would-KO features
  remain in the primary pipeline.
- Reconciled the current FP32 contract with legacy `bc_best_torch_fp16.pt`
  filenames and historical FP16 conversion examples.
- Updated the current-state ledger's Wikifita anchor and contradiction ledger.

### Pages updated

- [[pokemon_tcg_bc_curriculum_ablation]], [[pokemon_tcg_evaluation_results]],
  [[pokemon_tcg_training_pipeline]], [[pokemon_tcg_decision_trail]],
  [[pokemon_tcg_tournament_system]], [[pokemon_tcg_torch_inference]]
- [[pokemon_tcg_ai_battle]], [[pokemon_tcg_research_backlog]],
  [[pokemon_tcg_prospective_v2]], [[pokemon_tcg_current_state_reconciliation]]
- `memorias/projetos/pokemon_tcg/README.md`, `memorias/MEMORY.md`

### Evidence boundary

The older ablation remains a valid dated cohort with its own denominator and
runtime contract. The coherence change does not erase it or the later
RoPE-ND/MoE/vehicle/Apex roadmap. It only prevents historical experiment
interpretations and removed sidecar mechanics from being read as current
implementation claims. No project source, model or database was modified.

## [2026-08-15] reconcile + deepen | Pokémon TCG - curriculum lineage, brain artifacts and Metanoia provenance

Continued the sequential coherence pass by cross-reading the verified project
tree, the three August 14 Fita Labs PDFs, the Antigravity brain artifacts and
the main transcript's final decision checkpoints. The new layer separates
stage labels from cohort keys, records the 871-game, 420-game, 3,600-game and
teacher-sweep families, and anchors the Stage 3 incident and later plans to
their repository commits.

### Pages created

- [[pokemon_tcg_curriculum_stage_lineage]] - Stage 1–5 chronology, checkpoint/loss incident, cohort boundaries and provenance tuple for future experiments.
- [[pokemon_tcg_fitalabs_source_set]] - Source-role map for the Codex handoff v1.2, analytical thesis v0.5 and TBPTT/J-space dossier v1.1.

### Pages updated

- [[pokemon_tcg_cross_harness_provenance]] - main-brain artifacts (`master_project_state_and_handoff_plan.md`, `prompt_draft.md`, `schema_mermaid.md`, timezone derivation), quota stop and harness-transition boundary.
- [[pokemon_tcg_orchestration_artifact_ledger]] - Locked Meta roadmap claim class and full M2 oracle transcript identifier.
- [[pokemon_tcg_sqlite_schema_current]] - schema diagram artifact distinguished from current population and FK evidence.
- [[pokemon_tcg_metanoia_suite]] - commit-level provenance for Specs 01–06 and the RFC registration.
- [[pokemon_tcg_would_ko_prospective_search]], [[pokemon_tcg_repository_timeline]] - historical prospective/RoPE-ND wording made explicitly time-bounded after the August 3 sidecar removal.
- [[pokemon_tcg_docs_corpus_provenance]], [[pokemon_tcg_ai_battle]], [[pokemon_tcg_current_state_reconciliation]], [[pokemon_tcg_aug14_architecture_and_handoff_audit]], [[pokemon_tcg_handoff_acceptance_reconciliation]], `index.md`, `memorias/MEMORY.md`, `memorias/projetos/pokemon_tcg/README.md` - integrated the new lineage node, PDF source set and cross-links.

### Evidence boundary

The main transcript confirms the user's final August 14 instructions: stop the
teamwork preview at the quota boundary, preserve artifacts and future phases,
document the harness-specific channel protocol, and leave a recoverable handoff
for another harness. Those are provenance and authorization facts, not proof
that the Locked Meta roadmap, RoPEND/MoE modules, parity gate or 500-match
acceptance criteria were achieved. No project source, model or database was
modified.

## [2026-08-15] extend + cross-link | Pokémon TCG - invariant Elo, cross-project protocol and orchestration artifacts

Continued the sequential corpus integration by reconciling the stale
submission/Elo entry with the current `ResultsDB` implementation and the
read-only database snapshot. The update records the clipped Bradley-Terry
inversion, MD10 shrinkage, local-to-remote translation, draw-handling
difference, current deck observations and the Git lineage from source
separation through the PageRank/Abelian monograph.

### Pages created

- [[pokemon_tcg_orchestration_artifact_ledger]] - `.agents` topology, explorer handoffs, milestone state, missing handoff negative space and pending gates.
- [[elo-as-experimental-protocol]] - methodological comparison of Pokémon TCG, Antigravity, Co-Scientist and Red Team Arena Elo-like systems without merging their scales.

### Pages updated

- [[pokemon_tcg_submissions_and_elo]] - current database anchor, invariant-deck calculation, draw caveat, rating-layer boundary and commit provenance.
- [[pokemon_tcg_elo_pagerank_invariance]] - implementation boundary and cross-project Elo distinctions.
- [[pokemon_tcg_handoff_acceptance_reconciliation]], [[pokemon_tcg_cross_harness_provenance]], `index.md`, `memorias/MEMORY.md`, `memorias/projetos/pokemon_tcg/README.md` - added cross-links and durable reading-map entries.

### Evidence boundary

The project remains `develop` at `20d7d0d`; the `.agents` files remain
orchestration artifacts, not proof of implementation or passing tests. The
new rating prose is tied to source code, database observations and dated Git
commits. No Pokémon TCG code, model, database or future-stage document was
modified or deleted.

## [2026-08-15] audit + reconcile | Pokémon TCG — handoff claims and acceptance gates

Added a focused reconciliation for the normative and acceptance layers that surround the August 14 source handoff. The page distinguishes `AGENTS.md`, `GEMINI.md`, `PROJECT.md`, `TEST_INFRA.md`, `.agents/ORIGINAL_REQUEST.md` and the master RFC, then compares their milestones and acceptance criteria with the live source tree, current tests, database audit and Wikifita audit. It records the E2E reference/fallback boundary and preserves future modules and unchecked criteria without rewriting the project documents.

### Pages created or updated

- [[pokemon_tcg_handoff_acceptance_reconciliation]] - source roles, M1-M5 status, acceptance table, absent module paths, test-inventory limitation and Antigravity provenance.
- [[pokemon_tcg_mlx_migration]] - historical FP16 wording explicitly separated from the current strict-FP32 source contract.
- [[pokemon_tcg_ai_battle]], [[pokemon_tcg_docs_corpus_provenance]], `index.md`, `memorias/MEMORY.md`, `memorias/projetos/pokemon_tcg/README.md` - added the handoff reconciliation to the project graph.

### Evidence boundary

The current project remains `develop` at `20d7d0d`. The current source tree has no tracked RoPEND, MoE, vehicle-draft or Apex modules; `tests/e2e/` contains only the Tier 1 file and its reference fallback machinery; the direct database audit retains 2,946,336 FK violations; and the current Yan #633 Stage 4 slice is 27.9% over 140 games. The Wikifita double audit passes. No project code, tests or database were changed.

## [2026-08-15] expand + reconcile | Pokémon TCG — empirical monograph, augmentation contract and complete timeline

Continued the requested sequential integration from the verified project history and `docs/` corpus. This pass fills a provenance gap rather than replacing the previous audit: it adds a dedicated empirical-monograph page, a dedicated data-augmentation and vehicle-draft contract, and the missing August 8–13 curriculum, Stage 3, manuscript, rating, schema, governance and ETL history to the repository timeline. The later RoPEND/MoE, vehicle-draft and Apex material remains preserved as future lineage.

### Pages created

- [[pokemon_tcg_empirical_ablation_monograph]] - source-bound reading of the 420-game matrix, validation versus win-rate decoupling, deck saliency, historical FP16 collapse and Abelian rating proposal.
- [[pokemon_tcg_data_augmentation_vehicle_draft]] - separate contract for current derived supervision versus future synthetic 60-card vehicle drafting, with legality, cohort, reset and provenance fields.

### Pages updated

- [[pokemon_tcg_docs_corpus_provenance]] - corrected the inventory from 33 to 35 Markdown files, added the empirical ablation monograph and a source-to-page integration map.
- [[pokemon_tcg_repository_timeline]] - added the previously underrepresented August 8–13 research/governance phase and grouped low-signal commits without erasing them.
- [[pokemon_tcg_ai_battle]], `index.md`, `memorias/MEMORY.md`, `memorias/projetos/pokemon_tcg/README.md` - extended the durable reading map and current corpus count.

### Evidence boundary

The source repository remains `develop` at `20d7d0d` with 144 reachable commits. This pass used Git history, the 35-file `docs/` inventory, current Wikifita pages and the prior direct database/source reconciliation. No Pokémon TCG source code or database was modified. No future-stage material was deleted or promoted to implemented status.

## [2026-08-15] reconcile + refine | Pokémon TCG - current source, database and provenance coherence

Performed a second sequential integration pass after the corpus expansion. The
purpose was to make the entry points agree with the verified project endpoint,
not to delete dated experiments or future architecture. Project `develop` is
at `20d7d0d` with 144 reachable commits; Wikifita `main` started this pass at
`560ca3c`.

### Pages created or updated

- [[pokemon_tcg_current_state_reconciliation]] - current implementation matrix, database revalidation, milestone gates, post-blueprint Git lineage and contradiction ledger.
- [[pokemon_tcg_docs_corpus_provenance]] - extended the documentation timeline through `f988eb0`, `ea9317a`, `dbf298a`, `14ec607` and `20d7d0d`.
- [[pokemon_tcg_sqlite_schema_current]] - added the August 15 read-only `ResultsDB` revalidation, invariant Elo observations and latest tournament-ledger boundary.
- [[pokemon_tcg_aug14_architecture_and_handoff_audit]] - linked the current reconciliation to the August 14 source-versus-blueprint audit.
- [[pokemon_tcg_ai_battle]], [[pokemon_tcg_agent_architecture]], [[pokemon_tcg_training_pipeline]], [[pokemon_tcg_torch_inference]], [[pokemon_tcg_temporal_learning]], [[pokemon_tcg_train_config_reference]] - corrected stale current-state wording while preserving historical FP16, one-pass `topk` and discarded-scratch records as dated evidence.
- [[pokemon_tcg_bc_curriculum_ablation]], [[pokemon_tcg_training_overhaul_2026_07_29]] - labeled historical FP16 cohorts and linked the current strict-FP32 correction.
- `index.md`, `memorias/MEMORY.md`, `memorias/projetos/pokemon_tcg/README.md` - repaired the project reading map and durable current-state summary.

### Evidence boundary

The read-only database probe reproduced 139,783 matches, 3,065,706 match
steps, 30 datasets, 128 tournament summaries and 2,946,336 previously observed
foreign-key violations. It also reproduced deck 633 at 412.21 invariant Elo
and deck 251 at 577.39 over 420 local games each, with MD10 complete. These
observations do not establish whether tournament-local rows were reset between
runs. No project code, model or database was modified.

The current source contract is strict FP32 training and inference with
autoregressive multi-select and carried scratch memory. RoPE-ND, MoE, vehicle
drafting, Apex and later analyses remain documented future or experimental
material. The older 147-commit count remains as a dated historical snapshot;
the direct current count is 144.

## [2026-08-15] ingest + expand | Pokémon TCG — full docs corpus, Stage 3 postmortem and research lineage

Expanded the August 14 three-page audit into a source-mapped documentation layer covering the complete 33-file `docs/` corpus. The integration was performed sequentially: Git provenance and document roles were reconstructed first, then the Stage 3/4 training incident, neural engine, future architecture, ETL, manuscripts, cross-harness evidence, local platform, rating analysis, sampling analysis and Metanoia suite were added with explicit evidence boundaries.

### Pages created

- [[pokemon_tcg_docs_corpus_provenance]] — complete inventory, Git introduction timeline and source-role map.
- [[pokemon_tcg_stage3_training_failure_postmortem]] — Stage 3 data-window, validation, sparse auxiliary-loss and catastrophic-forgetting postmortem.
- [[pokemon_tcg_neural_engine_tokenization]] — as-built streams, action pointers, scratch/TBPTT and target delta.
- [[pokemon_tcg_ropend_moe_blueprint]] — RoPEND, stochastic Elo, Pilot/Vehicle MoE, 60-card data augmentation and Apex boundary.
- [[pokemon_tcg_data_oracle_etl]] — replay compiler, would-KO oracle, Parquet, zero-trust ETL and synthetic-data contract.
- [[pokemon_tcg_local_platform_design]] — current arena, target schema, replay and future worker/application boundary.
- [[pokemon_tcg_manuscript_and_empirical_synthesis]] — nine manuscript chapters and empirical/rating bridge.
- [[pokemon_tcg_cross_harness_provenance]] — Antigravity session/report ledger and channel/tokenizer adaptation.
- [[pokemon_tcg_elo_pagerank_invariance]] — Abelian rating, sampled graph and PageRank analogy.
- [[pokemon_tcg_sampling_timezone_identity]] — coverage, archive-window inference and ghost-team normalization.
- [[pokemon_tcg_metanoia_suite]] plus six linked specification pages — protocol, rules, adherence, tensor scaling, HALT and pedagogy.

### Pages updated

- [[pokemon_tcg_ai_battle]], `index.md`, `memorias/MEMORY.md`, `memorias/projetos/pokemon_tcg/README.md`

### Editorial boundary

The Stage 3 mechanism is documented as direct observation, code-validated history, source-reported handoff explanation and causal hypothesis in separate columns. The user's report of signal mixing and catastrophic forgetting is preserved as a serious working interpretation, while raw transcript search did not prove a specific row-level leak. RoPE-ND, MoE, vehicle drafting, Apex and later analyses remain visible as future or experimental work. No project source code, model or database was modified.

## [2026-08-14] audit + update | Pokémon TCG - source, ETL, ablations and tournament cross-check

Updated the Pokémon TCG knowledge graph from the August 14 repository handoff, live source tree, current ZIP and Parquet files, read-only SQLite snapshot, checkpoints and training logs, Git history, and the relevant Antigravity transcript provenance. The prior August 7 pages remain as dated historical snapshots.

### Pages created

- [[pokemon_tcg_aug14_architecture_and_handoff_audit]] - live implementation versus RoPE-ND/MoE and other future blueprints, with Antigravity handoff status and provenance IDs.
- [[pokemon_tcg_aug14_data_etl_database_audit]] - current replay, Parquet, SQLite, day metadata, auxiliary target and foreign-key observations.
- [[pokemon_tcg_aug14_ablations_tournaments]] - FP32 Stage 1-4 matrix, deck-conditioned Stage 4 result, curriculum validation logs and tournament persistence uncertainty.

### Pages updated

- [[pokemon_tcg_ai_battle]], [[pokemon_tcg_agent_architecture]], [[pokemon_tcg_training_pipeline]], [[pokemon_tcg_sqlite_schema_current]], [[pokemon_tcg_local_platform_status]]
- `memorias/projetos/pokemon_tcg/README.md`, `memorias/MEMORY.md`, `index.md`

### Evidence boundary

- Current direct reads: 30 replay ZIPs, 30 90-column Parquet partitions with 24,177,852 rows, SQLite Schema 2.0.0 with 139,783 matches and 128 tournament summaries.
- Preserved discrepancies: direct ZIP re-read counted 138,138 JSON members while the Antigravity survey recorded 140,511; day metadata and imported counts diverge on several dates.
- Preserved integrity anomaly: 2,946,336 foreign-key violations were observed; no purge or rebuild was run.
- Preserved future work: RoPE-ND, MoE, vehicle draft, Apex mode and later research phases were not deleted or relabeled as completed implementation.

The update documents current truth and uncertainty. It does not modify the Pokémon TCG source code or database.

## [2026-08-07] fix | Pokémon TCG — `--top-elo` filter page: SQL query and fail-mode corrections

Source-driven correction of [[pokemon_tcg_top_elo_curriculum_filter]] after a direct trace of `--top-elo` from CLI to episode mask in `scripts/bc/bc_train_mlx.py`. The original page was directionally right but had two hard implementation divergences:

- **Fail mode inverted.** Prior text said the filter "falls open" when a day has no `agent_elo_daily source='remote'` coverage. The code does the opposite: `allowed=None → continue → keep[i]=False` (line 2011-2013), i.e. every episode of that day is dropped. Corrected to explicit fail-closed statement, plus the second fail-closed path when the Parquet lacks `player_name`/`opponent_name` columns.
- **SQL query stylized as per-day loop.** Prior text showed one query per day; the actual implementation is a single global windowed query with `ROW_NUMBER() OVER (PARTITION BY aed.day_id ORDER BY aed.elo DESC)` (line 1995-2004). Replaced with the exact query.

Also added an eight-step "Implementation trace" (argparse → normalization → column scan → per-episode canonicalization → SQL scan → mask → log line → checkpoint provenance), a "Name provenance" block anchoring `player_name`/`opponent_name` to the builder's `_META_STR`, and a "What the filter does not touch" section that explicitly separates the training-time `--top-elo` from the evaluation-time `--sweep-source` (both hit the same SQLite but touch different tables and phases). Removed unsupported per-N concentration estimates (15–25%, 1–3%) in favour of the trainer's authoritative `kept K/T` log line.

Sources: live code trace at `scripts/bc/bc_train_mlx.py:1608-2023, 3772` and `scripts/bc/build_bc_from_zips.py:262, 404-405`.

## [2026-08-07] ingest + integrate | Pokémon TCG — canonical current-state cluster (training, inference, data, cache, tournament, config)

Comprehensive current-state pass over the Pokémon TCG cluster following the completed BC curriculum ablation. Nine new pages seven created and two existing pages updated so the wiki reflects the live pipeline instead of the pre-Parquet, pre-hierarchical-cache, pre-Round-Robin state.

### Sources and boundary

- Live code at `~/workdir/pokemon-tcg`: `scripts/bc/bc_train_mlx.py`, `scripts/tournament.py`, `rl/train_config.py`, `configs/train_config.schema.json`, `configs/train_config.json`, `rl/policy_mlx.py`, `rl/policy_infer_torch.py`, `agent/main.py`
- Live SQLite at `model/results.db` (schema_version 2.0.0) — sampled populated counts and enum values
- Live Parquet files under `data/bc_data/` — sampled 90-column schema and row-group layout
- Live tensorboard event files under `runs/` — extracted the 41-tag scalar set
- Ablation JSON reports at `experiments/bc_curriculum_suite/reports/`
- Repository commits: `fa38caa` (hierarchical KV cache + val streaming), `bb72620` (BC curriculum ablation suite + tournament JSON reports), `e8d6a4a` (SSD spill cleanup), `853ab34` (BSD-seq round-robin fix), plus 2026-08-07 config cleanup landing on `develop` in the pokemon-tcg repo

### Pages created

- [[pokemon_tcg_glossary]] — canonical vocabulary (TBPTT, KV cache, opt/gstep, chunk, sweep, source, aux heads, scratch registers)
- [[pokemon_tcg_parquet_dataset]] — 90-column day-partitioned Parquet format, row-group I/O unit, streaming reads, aux target contract, manifest sidecar
- [[pokemon_tcg_sqlite_schema_current]] — schema_version 2.0.0 with populated counts (551 agents, 552 decks, 22 days, 98,175 matches, 2.34M steps, 19 dataset rows, 6,552 daily agent-Elo rows) and the source enum
- [[pokemon_tcg_kv_cache_hierarchical]] — hot / transient / SSD spill three-tier row-group cache, pressure-driven eviction at 85%, promotion at ≥2 hits, `in_opt_step()` reentrant eviction suppression
- [[pokemon_tcg_train_config_reference]] — complete field-by-field and flag-by-flag reference, current baseline, deprecations (removed `--prefetch`, `--slab-rows`, `model_dir`)
- [[pokemon_tcg_training_pipeline]] — canonical end-to-end pipeline (MLX FP16, Muon+AdamW routing, TBPTT, aux heads, resumable checkpoints, 41-scalar tensorboard set)
- [[pokemon_tcg_tournament_system]] — tournament.py CLI (`--sweep-source`, `--opponent` multi-value, `--skip-baselines`, `--report-json`), sweep semantics, intra-suite round-robin as a step toward self-play
- [[pokemon_tcg_top_elo_curriculum_filter]] — `--top-elo N` exact query, coverage cost, findings (never wins on vs-publics; closes lb526 gap only at 5d_10ep_ON)
- [[pokemon_tcg_torch_inference]] — arena-side PyTorch FP16, checkpoint discovery, MLX→PyTorch conversion, autoregressive multi-select, per-side tracker state

### Pages updated

- [[pokemon_tcg_agent_architecture]] — refreshed with current numbers (128/4/3-4/512, scratch 16-32, ~1.30M params, ~845k Muon + ~456k AdamW), embeddings by zone, aux head table, explicit "no RoPE / no RoPE-ND, positions via zone types + explicit gather" statement
- [[pokemon_tcg_mlx_migration]] — marked as historical migration record; live-state pointer added
- [[pokemon_tcg_data_pipeline]] — marked as historical NPY era; superseded pointer to [[pokemon_tcg_parquet_dataset]]
- [[pokemon_tcg_tbptt_training_contract]] — cross-links added to the current-pipeline pages
- [[pokemon_tcg_physical_schema]] — pointer to the live-snapshot page ([[pokemon_tcg_sqlite_schema_current]])

### Integration

- `index.md` — Pokémon TCG cluster now lists the nine new pages
- `memorias/projetos/pokemon_tcg/README.md` — canonical reading map updated, status now `documented-training-pipeline-canonical-arch-validated-data-limited`

### Repository timeline reconciliation

- [[pokemon_tcg_repository_timeline]] — created a semantic timeline of the 147-commit `pokemon-tcg` repo, grouped by migration phase (A–F), then post-plan directions (prospective sidecar → removed; Parquet + aux + meta; hierarchical KV cache; BC curriculum ablation). Each phase links to the live-state wiki page. Includes a per-page evidence table mapping every current-state wiki page to the anchor commits that produced it, and a note explaining the many adjacent duplicate commits from a legacy `mlx-port` branch merge.

### Conversation-derived context integration

- [[pokemon_tcg_decision_trail]] — read the 205-message session transcript at `~/.claude/projects/-Users-alefita-workdir-pokemon-tcg/b99ca436-d9c3-4f23-ad6e-9ce870a00023.jsonl` and extracted the reasoning trail that produced today's pipeline. Chronological+thematic. Covers: the origin (900→300 Elo regression on the sidecar-augmented submission), the aux_prize_delta sign-inversion bug, `meta_lookup` business-rule-vs-exception distinction, the 32-hour epoch that Alefita killed manually, structured verb head disabled, latent b1/b2/b3 experiments retained as `_forward_b1`/`_forward_b2` hooks, the strategic time-budget decision to do the ablation instead of continuing the 32-hour run. Also captures the feedback rules (no fallbacks, business rules ≠ exceptions, no parallelism, TDAH/AHSD context, kaggle CLI available) that inform every future agent's operating envelope.
- [[pokemon_tcg_long_horizon_vision]] — Alefita's articulated multi-stage direction, extracted from her own words in the transcript: BC bootstrap → GRPO offline → frozen 900-Elo teacher as embedded world model + scratch zones for latent rollout → MoE specialists → two competition tracks. References the DeepSeek "Engram" and "thinking with visual tokens" papers she named. Documents which choices in the current design were made with this vision in mind (scratch registers scaling, meta buckets already zone-typed, aux heads architecturally identical to any future head).

### Historical pages cross-linked forward

Deepened the existing historical pages instead of leaving them as pointer stubs:

- [[pokemon_tcg_prospective_v2]] — new "What survived from this design" and "What died with the sidecar" sections. Explicit chain of ideas from the sidecar experiment into the current aux heads, meta buckets, and would-KO trio.
- [[pokemon_tcg_data_pipeline]] — new "Migration to Parquet (2026-08-03)" section with concrete NPY→Parquet correspondence table (episode_meta.npy → first-class columns, slab_rows → row group + KV cache, prospective sidecars → removed).
- [[pokemon_tcg_mlx_migration]] — "Delivery vs plan (2026-08-07 audit)" table mapping each A–F phase to landed commits and live-state pages, plus a section on post-plan work that stayed inside the "correctness before novelty" boundary the plan set.
- [[pokemon_tcg_would_ko_prospective_search]] — new "Status (audit 2026-08-07)" section reconciling the page's design claims against the live codebase: would-KO shipped and stable; RoPE-ND not implemented; MLX+PyTorch planner modules removed; autoregressive multi-select fallback is now the only inference path.

### Repository-side cleanup (pokemon-tcg)

### Repository-side cleanup (pokemon-tcg)

Removed truly deprecated fields from the training config so the schema matches the runtime:

- `--prefetch` / `prefetch` field removed from `TrainConfig`, JSON schema and CLI — the pyarrow Scanner reads ahead intrinsically and cross-batch retention is owned by the hierarchical KV cache
- `slab_rows` field removed from `TrainConfig` — the MLX trainer has no fixed-size mmapped slabs anymore
- `model_dir` field removed from `TrainConfig` and schema — nothing read it

## [2026-08-07] ingest | Pokémon TCG — BC Curriculum Ablation (10 configs + round-robin)

Recorded the outcome of a full behavioral-cloning ablation over `(1d, 3d, 5d) × (1ep, 10ep) × (top-Elo OFF, ON)` on the MLX trainer, followed by a no-sweep intra-suite round-robin between all 10 trained checkpoints. Findings: top-Elo curriculum never wins in matched pairs while the model is data-limited; coverage cost is visible (all ON zero vs `lb526` except `5d_10ep_ON`); anti-correlation between vs-publics and vs-peers rankings; architecture validated as data-limited, not architecture-limited. Current submission candidate: `5d_10ep_OFF` at 21.0% overall vs publics.

### Sources and boundary

- Live JSON reports under `~/workdir/pokemon-tcg/experiments/bc_curriculum_suite/reports/`
- Aggregated `results.md` written by the suite's phase-3 aggregator
- Full methodology and cross-references: [[pokemon_tcg_bc_curriculum_ablation]]

### Pages

- Created: [[pokemon_tcg_bc_curriculum_ablation]]
- Updated: [[pokemon_tcg_evaluation_results]] — added a section pointing to the new page; [[index]] — added the new entry to the Pokémon TCG cluster; `memorias/projetos/pokemon_tcg/README.md` — added canonical link and updated status.

## [2026-08-03] update | Canonical project routing and operational reporting

Recorded a cross-project workflow correction: a user-designated repository path
is authoritative and must be verified against live Git state before analysis or
writing. Added the same boundary to Codex and Wikifita instructions, together
with direct non-anthropomorphic operational reporting.

### Sources and boundary

- Alefita's explicit workflow instruction
- Live verification of `~/workdir/co-fita/redhat-clickfix-report`
- No files in that project were modified by this update

## [2026-08-01] research + directive | Human Principal, J-space, and Dialectical Method

Converted a live human-agent correction into an auditable governance protocol
and separated its operational, epistemological, and metaphorical claims.

### Pages Created

- [[human_principal_escalation]] — disclosure-before-remediation state machine,
  contact limits, identity compromise, catastrophic containment, and audit record
- [[j_space_inference]] — J-space as a relational constraint hypothesis with
  explicit evidence levels and falsifiable experiments
- [[dialectical_human_agent_method]] — Freirean dialogue, WALL-E directive
  governance, and the incident-to-protocol compiler
- [[associative_gadget_chain_memory]] — bounded ROP analogy for associative
  retrieval, temporal composition, and adaptive inference

### Pages Updated

- `CLAUDE.md`, `diretivas/index.md`, `index.md`, `memorias/MEMORY.md`
- [[communication_style]], [[identity_steering]]

### Boundary

The catastrophic scenario is documented only as an inert governance tabletop.
It does not authorize weapons operations, infrastructure interference, invented
channels, new authority, or transfer of harm to another population.

### Adversarial Refinement

- Separated authentication of claimant identity from authorization, scope, and
  policy constraints.
- Clarified that urgency never creates authority: only defensive safe-state
  actions enumerated and authorized in advance are eligible.
- Reframed the compiler and ROP language explicitly as bounded metaphors rather
  than implemented mechanisms.

## [2026-07-29] update | Pokemon TCG — TBPTT Training Contract

Documented the exact recurrent training accounting contract after reconciling
the source progress-bar task with the live MLX trainer.

### Pages Created

- [[pokemon_tcg_tbptt_training_contract]] — decision chunks, row budget,
  optimizer steps, scheduler phase, checkpoints, validation and progress
  accounting

### Pages Updated

- [[pokemon_tcg_ai_battle]], [[pokemon_tcg_temporal_learning]],
  [[pokemon_tcg_training_overhaul_2026_07_29]]
- `index.md`, `memorias/MEMORY.md`,
  `memorias/projetos/pokemon_tcg/README.md`

### Sources

- Live Pokemon TCG repository at commit `acbbea0`, especially
  `scripts/bc/bc_train_mlx.py`, `rl/train_config.py`,
  `configs/train_config.json`, `configs/smoke.json`, and
  `configs/train_config.schema.json`
- Source Codex task `019faa94-08c1-7d80-af02-873a75020b03`

## [2026-07-29] update | Pokemon TCG — Local Platform, Decks and Results

Separated current implementation, partial local-platform scope, deck strategy,
evaluation evidence and future research backlog into dedicated pages.

### Pages Created

- [[pokemon_tcg_local_platform_status]] — implemented, partial and planned
  boundaries for SQLite, tournament, dashboard, replay and ratings
- [[pokemon_tcg_deck_strategy]] — immutable submission `deck.csv`, local
  outer-loop deck selection and future `suggested_deck.csv` boundary
- [[pokemon_tcg_evaluation_results]] — historical local tournaments, one-epoch
  prospective tournament and ladder provenance rules
- [[pokemon_tcg_research_backlog]] — explicitly deferred research directions

### Pages Updated

- [[pokemon_tcg_ai_battle]], [[pokemon_tcg_local_overhaul]],
  [[pokemon_tcg_physical_schema]], [[pokemon_tcg_replay_data_model]],
  [[pokemon_tcg_dashboard_and_tournament]], [[pokemon_tcg_submissions_and_elo]],
  [[pokemon_tcg_experiments_and_anamnesis]], [[pokemon_tcg_ladder_and_research]]
- `index.md`, `memorias/MEMORY.md`, `memorias/projetos/pokemon_tcg/README.md`

### Sources

- Live Pokemon TCG repository at commit `acbbea0`, especially
  `rl/results_db.py`, `scripts/tournament.py`, `scripts/dashboard.py`,
  `scripts/rebuild_db.py`, `scripts/build_submission.py` and `TASK.md`
- Source Codex task `019fab09-3d12-7720-a0e3-48991bf3ec4c` for the July 29,
  2026 one-epoch tournament interpretation

## [2026-07-29] update | Pokemon TCG — Data Pipeline

Documented the replay ZIP to BC NPY to prospective sidecar pipeline as a
resumable, fingerprinted, smoke-bounded contract.

### Pages Created

- [[pokemon_tcg_data_pipeline]] — replay ZIP durability, BC NPY arrays,
  shard resume markers, base-stage idempotence, would-KO metadata,
  prospective sidecar validation, and daily sequence boundary

### Pages Updated

- [[pokemon_tcg_ai_battle]], [[pokemon_tcg_training_overhaul_2026_07_29]],
  [[pokemon_tcg_daily_corpus_training]]
- `index.md`, `memorias/MEMORY.md`, `memorias/projetos/pokemon_tcg/README.md`

### Sources

- Live Pokemon TCG repository at commit `acbbea0`, especially
  `scripts/bc/build_bc_from_zips.py`, `scripts/bc/build_bc_dataset.py`,
  `scripts/train_sequence.py`, and `configs/train_sequence.json`

## [2026-07-29] update | Pokemon TCG — Prospective V2 and Action Coverage

Expanded the Pokemon TCG knowledge graph with the current Prospective V2
implementation boundary. The sidecar/planner/runtime path is documented as
implemented scaffolding, while strict on-policy GRPO and ladder impact remain
future or unproven.

### Pages Created

- [[pokemon_tcg_prospective_v2]] — sidecar, lateral planner, group-relative
  objective, runtime fallback and strict-GRPO boundary
- [[pokemon_tcg_action_coverage]] — deterministic legal-action enumeration,
  bounded full-domain sampling and offline/runtime parity

### Pages Updated

- [[pokemon_tcg_ai_battle]], [[pokemon_tcg_would_ko_prospective_search]],
  [[pokemon_tcg_training_overhaul_2026_07_29]]
- `index.md`, `memorias/MEMORY.md`, `memorias/projetos/pokemon_tcg/README.md`

### Sources

- Live Pokemon TCG repository at commits `10c3750` and `acbbea0`
- Source Codex task `019fab09-3d12-7720-a0e3-48991bf3ec4c`, read through compact
  task summaries and local JSONL fallback

## [2026-07-27] update | Pokémon TCG — Local Overhaul and Post-Migration State

Updated the Pokémon TCG knowledge graph from the user-authorized review conversation and live project inspection. The MLX migration is recorded as user-reported complete with an eight-epoch base-model run and no Kaggle submission; the new local arena/schema/replay/dashboard design is explicitly marked as not yet implemented.

### Pages Created

- [[pokemon_tcg_local_overhaul]] — product boundary and decisions
- [[pokemon_tcg_physical_schema]] — relational schema, FKs and idempotency
- [[pokemon_tcg_replay_data_model]] — normalized replay and external visualizer
- [[pokemon_tcg_submissions_and_elo]] — model/deck/submission and rating layers
- [[pokemon_tcg_experiments_and_anamnesis]] — generic experiments and temporal history
- [[pokemon_tcg_dashboard_and_tournament]] — dashboard and synchronous arena

### Pages Updated

- [[pokemon_tcg_ai_battle]], [[pokemon_tcg_mlx_migration]], [[pokemon_tcg_ladder_and_research]]
- `memorias/projetos/pokemon_tcg/README.md`, `memorias/MEMORY.md`, `index.md`

### Sources and boundary

- User-authorized review conversation and transcript-derived decisions
- Live `~/workdir/pokemon-tcg` schema and replay/dashboard inspection
- Existing Wikifita project pages and editorial workflow

Future HashMath/on-chain compute-ledger ideas remain hypotheses outside the local implementation scope.

## [2026-07-25] feat | uv Ecosystem — Package Manager Best Practices

Dense reference page distilled from Pokemon TCG MLX migration project and global workflow learnings. Covers entrypoints, tool isolation, packaging hygiene, and anti-patterns.

### Pages Created

- [[uv-ecosystem]] — uv best practices: `uv run`, `[project.scripts]` entrypoints, `uv tool install`/`uvx`, `uv add` vs `uv pip install`, `uv sync`, PYTHONPATH anti-patterns, sys.path fixes, package structure checklist, subagent instructions, pyproject.toml template, checkpoint handling

### Key Rules Documented

- NEVER use `PYTHONPATH=.` — fix project structure instead
- Always use `[project.scripts]` entrypoints in pyproject.toml
- `uv tool install` for standalone CLI tools (isolated from project venv)
- `uvx` for one-off tool runs
- `uv add` for project deps, `uv pip install` for temporary only
- `uv sync` required after any pyproject.toml change
- Every importable directory needs `__init__.py`
- `sys.path.insert(0, ...)` is a code smell (exception: Kaggle sandbox)
- Subagents must always be instructed to use `uv run`

### Cross-References

- [[preferencias-tecnicas]] — The directive that mandates uv
- [[pokemon_tcg_mlx_migration]] — Project where these rules crystallized

---

## [2026-07-25] feat | Compute Optimization — MLX, OpenCL, Metal, Async Pipelines

Deep technical audit of GPU compute patterns across hashfita (Chronobreaker) and crustybike. 7 OKF pages + hub, ~2800 lines total. Source: exhaustive analysis of both projects' ML implementations, GPU kernels, async pipelines, and floating-point strategies.

### Pages Created

**compute-optimization/ (4 pages)**
- [[compute-optimization]] — Hub: cross-project GPU compute patterns, architecture map, key findings summary
- [[mlx-opencl-bridge]] — Buffer transfer chains, 3-copy problem, zero-copy via mx.take(), shared memory opportunities
- [[opencl-kernel-design]] — File isolation vs embedding, dynamic loading, 2D NDRange, atomic early exit, midstate caching, Metal port conventions
- [[gpu-async-pipelines]] — run_in_executor, 10-worker prefetch ("Tropa de Choque"), starvation prevention, TDR recovery, auto-tuning
- [[floating-point-determinism]] — float32 exclusivity, deterministic int64 quantization, PHI as rational, bfloat16 opportunity analysis

**compute-optimization/ml-architectures/ (2 pages)**
- [[ring-transformer]] — Custom MLX transformer: 17-ring input, 4D Continuous RoPE, cyclic loss, 3-pass training step, ~272k params
- [[muonclip-optimizer]] — Newton-Schulz orthogonalization, QK-Clip (tau=100), rank-1 matrix hazard, 1D variant for holographic weights

### Key Findings

- hashfita has a 3-copy chain (OpenCL→numpy→MLX) — biggest memory bottleneck
- crustybike achieved zero-copy via mx.take() for Merkle permutations
- Both use float32 exclusively — bfloat16 opportunity on M3 Pro (2× throughput, ½ memory)
- Metal kernels exist in crustybike but aren't loaded by Python code
- TDR recovery pattern in crustybike critical for 24/7 operation, missing in hashfita
- MuonClip discovered that Newton-Schulz destroys rank-1 matrices (use_muon_for_linears=False)
- Training step crosses OpenCL↔MLX framework boundary — architectural bottleneck

### Sources

- `~/workdir/hashfita/` — full project tree (CLAUDE.md, pyproject.toml, src/, workers/, proof_of_concept/)
- `~/workdir/crustybike/` — full project tree + git history
- Analysis artifact: `ml-optimization-analysis.html` (2026-07-25)

### Cross-References

- [[pokemon-tcg-mlx-migration]] — MLX migration patterns from Kaggle project
- [[antigravity-unified-scheduler]] — Double-SHA-256 optimization in scheduler context
- [[co-fita-hashmath-analysis]] — SHA-256 chain analysis, persistence gap
- [[preferencias-tecnicas]] — Hardware specs (M3 Pro, 24GB)

---

## [2026-07-25] docs | Pokémon TCG AI Battle project knowledge integrated

Integrated the current state and technical decisions of `~/workdir/pokemon-tcg` into the Kaggle knowledge area. This records the Mikaelzinho PyTorch behavioral-cloning snapshot, the MLX migration in progress, the current architecture ceiling, the temporal-learning contract and the research backlog. It does not create a task page and does not modify the project code.

### Pages created

- [[pokemon_tcg_ai_battle]] — project hub, source authority and evidence boundary
- [[pokemon_tcg_agent_architecture]] — current agent architecture and limitations
- [[pokemon_tcg_mlx_migration]] — MLX migration scope, P0 corrections and releases
- [[pokemon_tcg_temporal_learning]] — sequential metadata, persistent registers and TBPTT
- [[pokemon_tcg_ladder_and_research]] — ladder observations, Elo-oriented decisions and deferred research
- `memorias/projetos/pokemon_tcg/README.md` — durable project memory index

### Editorial boundary

The original author's PPO/self-play implementation is treated as historical context only. The Mikaelzinho PyTorch behavioral-cloning implementation is the migration reference. The MLX port is documented as in development. Future ideas are explicitly marked as hypotheses or backlog, not current capabilities or decisions.

### Sources

- `~/workdir/pokemon-tcg/CLAUDE.md` and `TASK.md`
- `rl/policy.py`, `rl/policy_mlx.py`, `rl/encoder/encoding.py`, `agent/main.py`
- `scripts/bc/bc_train.py`, `scripts/bc/bc_train_mlx.py`
- `model/eval_results.txt`, `result.html` and repository metadata
- originating technical discussion, reconciled against repository evidence

## [2026-07-21] feat | Deep Expansion — 28 New Pages, ~13,000 Lines

Phase 2 deep expansion: 6 research reports, 28 specialized subpages.

### sdk-test/ (7 new)
- [[sdk-test-localharness]], [[sdk-test-protobuf]], [[sdk-test-hooks-deep]], [[sdk-test-auth-chain]], [[sdk-test-grpc-protocol]], [[sdk-test-decompiled-patterns]], [[sdk-test-deepseek-heartbeat]]

### co-fita/ (7 new)
- [[co-fita-dynamic-workflows]], [[co-fita-sdlc-protocol]], [[co-fita-green-mile]], [[co-fita-red-team-blue-team]], [[co-fita-strawberry-research]], [[co-fita-hashmath-analysis]], [[co-fita-honcho-comparison]]

### wikifita-site-architecture/ (6 new)
- [[wikifita-site-pagerank]], [[wikifita-site-graph-construction]], [[wikifita-site-search-system]], [[wikifita-site-rendering-pipeline]], [[wikifita-site-corpus-pipeline]], [[wikifita-site-deployment]]

### sprite-mobile/ (3 new)
- [[sprite-mobile-knocklace]], [[sprite-mobile-code-review]], [[sprite-mobile-redux]]

### antigravity-ecosystem/ (5 new)
- [[antigravity-hermes-wiki]], [[antigravity-strawberry-research]], [[antigravity-unified-scheduler]], [[antigravity-browser-prompting]], [[antigravity-mcp-tools]]

### Security
- Credentials redacted from sdk-test/RESEARCH_REPORT.md

---

## [2026-07-20] feat | BairesDev AppSec Playbook

- Added a multi-page private editorial sourced from BairesDev's public blog.
- Covered company delivery worldview, DevSecOps, secure SDLC, SAST/DAST/IAST/SCA, threat modeling, supply chain, cloud, AI-assisted development, governance and metrics.
- Added cross-links and source boundaries distinguishing published claims from durable editorial synthesis.

## [2026-07-20] refactor | BairesDev AppSec Playbook made durable

- Removed interview-specific language, Morgan Stanley context, dated preparation framing and likely-question material from the durable playbook.
- Renamed the governance page to `governance-metrics.md`; process-specific interview notes remain isolated in the separate opportunity project.

## [2026-07-20] feat | AppSec methodology consolidated from private research

- Added a durable methodology page connecting RD Saúde, Security by Design, API governance, offensive testing, vulnerability management and SDLC controls.
- Sources included Wikifita career/CAMDOM records and private Drive research artifacts; no secrets were copied.

## [2026-07-20] feat | BairesDev Morgan Stanley AppSec interview

- Recorded Clara's economic proposal and the 2026-07-21 technical interview context on the private `personal` branch.
- Added the unresolved Erick interviewer record and preparation notes for secure SDLC, SAST, DAST, SCA, IAST and AI security automation.
- Confirmed Morgan Stanley from the calendar invitation; did not promote public-search homonyms to a verified identity.
- Correction: Erick is the BairesDev interviewer; 2026-07-21 is the confirmed date. Targeted BairesDev blog-author search found no matching Erick/Eric byline for AI, AppSec, DevSecOps, SDLC, SAST or DAST.
## [2026-07-21] feat | Co-Fita Hyper-Harness — 15 Pages, 5054 Lines

Full integration of the co-fita project. 4 parallel subagents. Source: harness_core, dashboard, unified-harness-wiki, .deli, hyper_workspace, .agents/skills.

### 🧬 co-fita/ (15 pages)

**Core System**
- [[co-fita-harness]] — Hub: 3 pillars, architecture, version history
- [[co-fita-orchestrator]] — 5-phase loop: Ideation → Execution → Audit → Fixation → Governance
- [[co-fita-components]] — 10 components deep-dive
- [[co-fita-governance]] — Democratic Centralism, proposals, congress, chancela

**Dashboard**
- [[co-fita-odysseus]] — FastAPI + Vanilla JS, 28 models, 54 routes
- [[co-fita-mcp-servers]] — 4 MCP servers: email, memory, RAG, image gen
- [[co-fita-kanban]] — 7-column governance kanban

**Research & Knowledge**
- [[co-fita-harness-wiki]] — 15+ source analyses, intellectual foundation
- [[co-fita-deli-governance]] — Quinquennial plans, strawberry research
- [[co-fita-hyper-workspace]] — Live state, failure tree, orchestrator logs
- [[co-fita-agent-skills]] — Telegram, web research pipeline, AutoResearch

**Infrastructure & Security**
- [[co-fita-clickfix-vaccine]] — Vaccine branch: detect/vaccinate/report
- [[co-fita-pam-setup]] — PAM Touch ID macOS
- [[co-fita-infrastructure]] — Docker: SearXNG, ChromaDB, vLLM-MLX
- [[co-fita-research-operations]] — Research ops methodology

---

## [2026-07-21] feat | Sprite Mobile + Wikifita Site + Antigravity Ecosystem — 9 Pages

Integration of 3 additional projects.

### 📱 sprite-mobile/ (4 pages)
- [[sprite-mobile]] — Knocklace Collection RN 0.74, Skia, BLE, Redux
- [[sprite-mobile-architecture]] — Tech architecture deep-dive
- [[sprite-mobile-components]] — Components and screens catalog
- [[sprite-mobile-ble]] — BLE integration, CAMDOM comparison

### 🌐 wikifita-site-architecture/ (3 pages)
- [[wikifita-site-architecture]] — Next.js 16 + Vinext + Cloudflare Workers
- [[wikifita-site-pipeline]] — 7-phase wiki pipeline with PageRank
- [[wikifita-site-llm-integration]] — LLM endpoints for agent consumption

### 🧠 antigravity-ecosystem/ (2 pages)
- [[antigravity-ecosystem]] — Antigravity 2.0 ecosystem hub
- [[antigravity-brain-artifacts]] — Brain artifact analysis

---

## [2026-07-21] feat | sdk-test — 10 Pages, 4184 Lines

Full integration of the probe/sdk-test project. 3 parallel subagents. Source: Antigravity SDK, Go binary RE, internal wiki, agent rules.

### 🔍 sdk-test/ (10 pages)
- [[sdk-test-antigravity-sdk]] — Python SDK: Agent/Conversation/Connection, hooks, triggers, policies
- [[sdk-test-agent-api]] — API interception: wrapper shim, data capture, JSONL logging
- [[sdk-test-reverse-engineering]] — 3305 decompiled functions, protobuf schemas, credentials
- [[sdk-test-architecture]] — Jetsky, auth chain, token lifecycle, gRPC protocol
- [[sdk-test-integrations]] — DeepSeek V4, heartbeat, Gemma 4 local, SDK integration
- [[sdk-test-synthesis]] — Epiphany synthesis, Gnosis thesis, RFC-001 auto token
- [[sdk-test-agent-rules]] — 10 rules: epistemic search, channel protocol, steganography
- [[sdk-test-preprocessor]] — PreProcessor plugin: TCG workflow, Unicode tokens
- [[sdk-test-tools]] — Decompilation pipeline: Ghidra batch, Go pclntab
- [[sdk-test-samurai-analogy]] — Samurai analogy reformulation, AI identity, hyperstition

### Cross-references
- RE trilogy: [[bezetacil-toolset]] ↔ [[sdk-test-reverse-engineering]] ↔ [[claude-desktop-bundles]]
- SDK: antigravity-2.0 ↔ [[sdk-test-antigravity-sdk]] ↔ [[mcp-ecosystem]]
- Protocol: [[unit-distance-channel-protocol]] ↔ [[sdk-test-agent-rules]] ↔ [[sdk-test-preprocessor]]
- Identity: [[unit-distance-philosophy]] ↔ [[unit-distance-steganographic-cot]] ↔ [[sdk-test-samurai-analogy]]

---

## [2026-07-21] feat | Co-Scientist + Red Team Arena — 17 Pages, 5317 Lines

Full integration of two research projects into wikifita. 4 parallel subagents.

### 🧬 co-scientist/ (8 pages)
- [[co-scientist]] — Hub: Google DeepMind Co-Scientist re-implementation
- [[co-scientist-agents]] — 7 agents: Supervisor + Generation, Reflection, Ranking, Evolution, Proximity, Meta-review
- [[co-scientist-prompts]] — 14 Jinja2 templates, structured output via tool use
- [[co-scientist-pipeline]] — Full pipeline: goal → generation → reflection → ranking → evolution → meta-review
- [[co-scientist-infrastructure]] — 9 LLM providers, FAISS, SQLite 15 tables, token budgets
- [[co-scientist-elo-tournament]] — Elo for model comparison, gold-set scoring
- [[co-scientist-evaluation]] — Bench system, AML drug repurposing, presets
- [[co-scientist-safety]] — Injection defense, safety classifier, citation verifier

### 🔴 red-team-arena/ (9 pages)
- [[red-team-arena]] — Hub: AI safety evaluation, attacker-defender-judge triangle
- [[red-team-arena-agents]] — 4 attacker types, 3 defenders, judge
- [[red-team-arena-elo]] — Elo K=32, dual pools, cross-project comparison
- [[red-team-arena-experiments]] — Transferability, category breakdown, adaptive vs static
- [[methodology]] — Blind judge, Wilson CIs, z-tests
- [[results]] — 953 rounds, 41.8% attack success
- [[code-quality]] — Architecture review, test coverage
- [[seed-prompts]] — 120 prompts, 8 categories
- [[red-team-arena-dashboard]] — Streamlit, Plotly charts

### Cross-references woven
- Elo trilogy: [[unit-distance-elo-ranking]] ↔ [[co-scientist-elo-tournament]] ↔ [[red-team-arena-elo]]
- Multi-agent: [[multi-agent-methodology]] ↔ [[co-scientist]] ↔ [[red-team-arena]]
- AI safety: [[kaggle-agent-security]] ↔ [[attack-primitives]] ↔ [[red-team-arena]]

### Index Updated
- `index.md` — 🧬 co-scientist/ (8 pages) + 🔴 red-team-arena/ (9 pages), 11 new tags

---

## [2026-07-21] feat | Unit Distance Research — 23 Pages, 6040 Lines

Full integration of the unit-distance-csfita research project into wikifita. 5 parallel subagents. Source: Antigravity 2.0 harness, 4 sessions, 16 hypotheses.

### Section: 🔬 unit-distance/ (23 pages)

**Mathematical Research (6 pages)**
- [[unit-distance]] — Hub overview, research journey, metrics
- [[unit-distance-problem]] — Erdős problem, OpenAI disproof, Sawin's bound
- [[unit-distance-hypotheses]] — Complete H1–H16 catalog with Elo scores
- [[unit-distance-h16-breakthrough]] — Degree 16 CM construction, δ=0.019603
- [[unit-distance-golod-shafarevich]] — GS inequality, Frattini quotient
- [[unit-distance-cm-fields]] — CM fields, class groups, genus theory

**Methodology & Swarm (4 pages)**
- [[unit-distance-methodology]] — Cognitive swarm, Generate→Debate→Evolve
- [[unit-distance-elo-ranking]] — Elo calibration 1800–2700+
- [[unit-distance-anti-contamination]] — Integrity protocol, H3 incident
- [[unit-distance-peer-review]] — Nature review, H10 rejection, H16 approval

**Tokenization & Protocol (4 pages)**
- [[unit-distance-channel-protocol]] — Channel protocol, HAL mapping
- [[unit-distance-tokenizer-analysis]] — Gemma 4 tokenizer, special tokens
- [[unit-distance-steganographic-cot]] — Steganographic CoT, persona steering
- [[unit-distance-grpo-inference]] — GRPO tournament inference

**Development Journey (4 pages)**
- [[unit-distance-development-journey]] — 4-session chronology
- [[unit-distance-seasons]] — Season 1–3 progression
- [[unit-distance-dead-ends]] — Failed hypotheses, lessons learned
- [[unit-distance-artifacts]] — 18 PEP 723 scripts inventory

**Philosophy & Ecosystem (4 pages)**
- [[unit-distance-philosophy]] — Poincaré cycles, hypersigils, metanoia
- [[unit-distance-ecosystem-handoff]] — Antigravity → Anthropic HAL
- [[unit-distance-creational-mythos]] — DeepSeek transcript analysis
- [[unit-distance-alefita-co-researcher]] — Alefita as co-researcher

### Metrics
- 23 OKF pages, 6040 lines total
- 5 parallel subagents (math, methodology, tokenization, journey, philosophy)
- Cross-references to existing wikifita: antigravity-2.0, [[multi-agent-methodology]], [[claude-mythos-card]], [[hypersigil-research]], [[kaggle-agent-security]], [[camdom-career-profile]], [[user_alefita]]

### Index Updated
- `index.md` — new 🔬 unit-distance/ section with 22 pages, 12 new tags

---

## [2026-07-20] feat | Editorial Workflow — Approved and Internalized

Full editorial front formalized as directive. Merged artifact-identity into wikifita skill.

### Directive Created
- [[editorial-workflow]] — 5-phase workflow, 12 validated rules, content taxonomy, failure mode taxonomy, quality gates

### Skill Updated
- wikifita skill (Cowork + Codex) — editorial workflow, content taxonomy, artifact visual identity, wikilink convention all integrated
- artifact-identity absorbed: visual identity rules now live in wikifita skill + `diretivas/identidade-visual.md`

### Index Updated
- `index.md` — new diretivas/ section with editorial-workflow, editorial tag added


---

## [2026-07-20] feat | CAMDOM Section Expanded — 13 Pages, 8390 Lines

Deep technical documentation of the entire CAMDOM codebase. Source: `condom-app/` (429 commits).

### Pages Created (8 new)
- [[camdom-skia-rendering]] — Ball component, ImageShader textures, Blur, ParagraphBuilder, dual-ball system, GPU rendering pipeline
- [[camdom-native-modules]] — Android GATT Server (Kotlin), iOS CBPeripheralManager (Swift), Expo Modules bridge, packet encoding/decoding, background execution
- [[camdom-gesture-system]] — Pan gesture, hold-to-disconnect (517ms), haptic feedback (9 points), retry mechanism, shouldBlock pattern, thread boundary map
- [[camdom-ui-state-machine]] — 7-state UI visualization, ball scaling (1x→3x), sound orchestration, SharedValue/DerivedValue registry (17+13)
- [[camdom-i18n-and-menu]] — i18n (pipe-delimited text, f: prefix), MMKV storage (8 keys), permissions, menu architecture, useLateralMenu hook
- [[camdom-onboarding]] — 2-step flow: HowTo → permissions, iOS vs Android behavior differences
- [[camdom-skia-components]] — 16 Skia symbols, Ball 4-layer rendering, ParagraphBuilder system, gesture-driven interaction
- [[camdom-animation-engine]] — Generator worklets, 6 primitives, comparison vs Reanimated/Moti, standalone lib proposal
- [[camdom-code-quality]] — Code review: 8967 LoC, god object (BleManager 1877 lines), type bug, zombie code, 0% tests, 13+ magic numbers
- [[camdom-system-design]] — Delivered vs conceptual: RSSI proximity disabled, shared state dilemmas, disconnect protocol (7 sub-states), 6 alarm paths, zombie features
- [[camdom-people]] — 3 organizations, 8 team profiles, communication architecture, trust model, pressure points
- [[camdom-development-timeline]] — May–Sep 2024 day-by-day chronology, version chronicle 2.17→3.3.0+
- [[camdom-store-review-saga]] — Apple/Google review: NFC, privacy, Bluetooth strings, CodePush, 30+ submission events
- [[camdom-debugging-stories]] — 13 bug narratives: "cache" bug, Android×Android, server election, bond state, OpenReplay
- [[camdom-scope-and-delivery]] — What shipped vs cut vs added, Felipe's philosophy, OTA strategy

Source: WhatsApp chats (_chat_felipe_almeida.txt + _chat_camdom.txt)

### Previously Created (5 pages)
- [[camdom]] — Executive overview, impact metrics, awards
- [[camdom-architecture]] — BLE mesh topology, native modules, state machine
- [[camdom-packet-protocol]] — Complete BLE packet reference, Mermaid diagrams
- [[camdom-stack-migrations]] — React Native, Expo SDK, JSI, TurboModules
- [[camdom-design-system]] — Palette, typography, animations, sound design

### Design System
- `diretivas/design-systems/camdom/index.md` — CAMDOM visual identity reference

---

## [2026-07-20] feat | CAMDOM Section — 5 Pages, 2946 Lines

Comprehensive CAMDOM documentation added to wikifita. Source: `condom-app/` (429 commits, 6-month solo development).

### Pages Created
- [[camdom]] — Executive overview: impact metrics (3.5B impressions, 850M people, 300K downloads, 27,400% ROI), partners, awards (29 international), development timeline
- [[camdom-architecture]] — BLE mesh topology, native module layer (Kotlin/Swift/TS), connection state machine (7 states), advertising system, service UUID, privacy-by-design
- [[camdom-packet-protocol]] — Complete BLE packet reference: 9 packet types, server/client handlers, RSSI proximity system, Mermaid sequence diagrams, 10 edge cases documented
- [[camdom-stack-migrations]] — React Native 0.74→0.82, Expo 51→54, JSI migration, TurboModules, Gradle/CocoaPods, patches, CI/CD evolution, dependency tables
- [[camdom-design-system]] — Color palette (#132037/#9DB0B9/#b2140c), BasicSans typography (14 variants), generator-based animation engine, 5 sound types, haptic language

### Design System Added
- `diretivas/design-systems/camdom/index.md` — CAMDOM design system reference alongside Anthropic

### Index Updated
- New section: 🛡️ camdom/ with 5 pages
- Tags added: camdom, ble, mobile, design-system

---

## [2026-07-17] feat | Branch Model, Site Pipeline, English Standardization

Major session: ClickFix integration, full English translation, branch model, site documentation.

### Branches Created
- `personal` — Private content (emails, salary negotiations, sensitive references). Not deployed.
- `main` — Public knowledge base + site content. Pushes update wikifita.aleffita.dev.

### Pages Created
- [[wikifita-site]] — Front-facing site pipeline: Next.js, branch model, public/private separation
- Sensitive contact response archived to `personal` branch

### Pages Translated
- All 47+ .md files translated from Portuguese to English (commits `e4178ba`, `2326760`)

### ClickFix Integration (previous commit)
- 8 wiki pages in `clickfix/` (1750 lines total)
- [[clickfix-handoff]], [[clickfix-attack-chain]], [[clickfix-bddr]], [[clickfix-c2-bot]], [[clickfix-stealers]], [[clickfix-iocs]], [[clickfix-incident-response]], [[clickfix-infrastructure]]

### Updates
- `CLAUDE.md` — Branch model and site pipeline documented
- `index.md` — wikifita-site added to infraestrutura section
- Wikifita skill — Updated to English, site pipeline documented, synced with codex

## [2026-07-15] update | Wikifita Contract — MDX and Media by URL

Canonical contract updated to prepare rich content while maintaining wiki idempotency.

- **Documents:** `.md` and `.mdx` accepted as canonical documents
- **Raw:** mandatory lossless preservation, no arbitrary reserialization
- **Media:** contractual support for remote references via secure URL; no `/assets`, no bucket, no Google Drive, no upload, and no ingest at this stage
- **Audit:** `wikifita_audit.py` now validates `.md`, `.mdx`, markdown/wiki links, media URLs, and symlinks `AGENTS.md -> CLAUDE.md`
- **Hooks:** new `scripts/install_hooks.py` installs/verifies `.githooks` idempotently in new environments
- **Atlas:** moodboard kept only as a temporary visual reference; consolidation into directives left for a separate stage

## [2026-07-15] ingest | Red Hat ClickFix APT — Threat Intelligence Study

Complete integration of forensic research from the `redhat-clickfix-report` project into wikifita. 8 new OKF pages created in `clickfix/`.

### Pages created
- [[clickfix-handoff]] — Executive overview, IOC summary, disclosure timeline
- [[clickfix-attack-chain]] — Domain hijack, ClickFix, LaunchAgent, fake PAM bypass (docs 01+02)
- [[clickfix-bddr]] — Blockchain Dead-Drop Resolver, ABI decoding, Python reproduction (docs 03+09)
- [[clickfix-c2-bot]] — Bot bmodule, 5 command types, tccutil reset All (doc 04)
- [[clickfix-stealers]] — smodule + lmodule, 60 wallet targets, LevelDB strategy (docs 05+06)
- [[clickfix-iocs]] — IOC inventory, 20 MITRE ATT&CK techniques, Suricata rules, remediation guide (doc 07)
- [[clickfix-incident-response]] — 160s containment, AI-assisted IR, refusal problem (doc 08)
- [[clickfix-infrastructure]] — Dated infrastructure observations, prospectively registered hypotheses and 16 monitoring records (doc 09 + monitoring/)

### Sources
- `redhat-clickfix-report/README.md` — executive report
- `redhat-clickfix-report/docs/01-09` — 9 technical deep-dives
- `redhat-clickfix-report/monitoring/` — 16 monitoring cycles (Jul 02–10)

### Updated indexes
- `index.md` — new section 🔴 clickfix/ with 8 pages, tags added (clickfix, threat-intelligence, blockchain)
- `memorias/projetos/redhat_clickfix_report/README.md` — project memory created (fixes pending reference)
- `memorias/MEMORY.md` — project reference updated

### Metrics
- 1750 lines of technical documentation
- 8 OKF pages with frontmatter, tags, cross-refs
- Coverage: attack vector → persistence → C2 → stealer → IOCs → IR → infrastructure → monitoring

## [2026-07-14] create | Professional Contact — BairesDev AppSec

Professional connection added. Details on `personal` branch only.

- **Status:** archived to `personal` branch (2026-07-21 security curation)

## [2026-07-14] fix | Versioned Hooks and Codex Compatibility

Replaced the informative post-commit hook with a pre-commit integrity flow and post-tree-change reconciliation.

- **Audit:** validates and, with `--fix`, creates relative symlinks `AGENTS.md -> CLAUDE.md`
- **Pre-commit:** complete and blocking audit
- **Post-checkout/post-merge:** idempotent symlink reconciliation
- **Hooks:** versioned in `.githooks/` and activated via `core.hooksPath`
- **Runtime:** `uv` only, no fallback to direct Python

## [2026-07-07] update | Visual Identity — Anthropic Dark Mode as Standard

Visual identity directive updated. Anthropic dark mode approved as standard after artifact pricing-analysis-v2.

- **Palette:** bg #171717, warm accent #d4a574, green #66bfa2, red #e8706a
- **Standard:** Source Serif 4 headings, Inter body, JetBrains Mono code
- **Light mode:** kept as an alternative for PDFs and official documents
- **Reference:** artifact pricing-analysis-v2 (pricing providers analysis)
- **Page:** [[identidade-visual]]

## [2026-07-07] feat | Skills, Hooks, and Scripts — Workflow Overhaul

Complete implementation of the workflow plan approved by Alefita.

### Skills
- **artifact-identity** — skill that triggers on artifact creation, injects Anthropic dark mode visual identity, metadata, typography. Never create artifacts in the project folder unless the user explicitly requests it.
- **wikifita** (revised) — updated skill. Removed query/ingest/lint. Now: contextual reading, proactive writing, OKF frontmatter, directives applied automatically.

### Scripts
- **wikifita_audit.py** — PEP 723 script, `uv run`. Validates OKF frontmatter (type, title, description, tags, timestamp), audits markdown links, verifies index coverage, detects deprecated refs. Zero dependencies.

### Git Hooks
- **pre-commit** — validates OKF frontmatter on all staged `.md` files. Blocks commit if required fields are missing. NEVER use `--no-verify`.
- **post-commit** — runs `uv run scripts/wikifita_audit.py` after each commit.

### Directives
- **preferencias-tecnicas.md** — new directive. uv as package manager, PEP 723 for scripts, pyproject.toml never requirements.txt, artifacts in outputs, hardware (M3 Pro, RTX 2060), mandatory gitflow.
- **identidade-visual.md** — updated with Anthropic dark mode as standard.
- **design-systems/anthropic/index.md** — updated with dark mode palette.

### CLAUDE.md
- Updated to reflect that wikifita is persistent memory, not a query/ingest wiki. Gitflow enforced via hooks.

### Pages
- [[preferencias-tecnicas]] — new directive
- [[identidade-visual]] — updated
- [[design-systems/anthropic]] — updated
- [[CLAUDE.md]] — updated

---

## [2026-07-06] ingest | FitaLabs Gateway Journey — LiteLLM Proxy + 1P Bundle

Complete documentation of the 4-phase migration:
1. FastAPI Custom → LiteLLM Router as library
2. LiteLLM Router → LiteLLM Proxy Server (main app)
3. PostgreSQL, Prisma, functional Admin UI
4. 1P Bundle hosted at `claude.unifita.app` (47 files, CDN fallback, Caddy auth injection)

**Pages**: [[fitalabs-gateway-1p-jornada]] (overview), [[litellm-proxy-migration]] (proxy migration),
[[fitalabs-gateway-modelos]] (models), [[fitalabs-gateway-caddy]] (Caddy),
[[fitalabs-gateway-bundle-1p]] (1P bundle), [[fitalabs-gateway-auth]] (auth),
[[fitalabs-gateway-docker]] (Docker)

**Key findings**:
- `CONFIG_FILE_PATH` is the env var the proxy reads (not `LITELLM_CONFIG_PATH`)
- Admin UI requires PostgreSQL (Prisma schema is `postgresql`, not `sqlite`)
- `libatomic1` required in Dockerfile for Prisma/Node binaries
- Caddy `header_up Authorization` is more reliable than Service Worker for auth injection
- SPA fallback with `header Accept *text/html*` avoids serving HTML for dynamic imports
- CDN fallback with `handle_errors 404` allows downloading chunks on-demand

**Mapped 1P SPA endpoints**:
- `GET /edge-api/bootstrap` — bootstrap with features + settings
- `GET/PUT /api/account_profile` — read and update settings
- `POST /api/accounts/me/organizations/get_or_create_chat_organization`
- `GET /api/auth/login_methods`, `POST /api/auth/send_magic_link`
- `POST /api/event_logging/v2/batch`, `GET /api/updates`

---

## [2026-06-30] fix | Remove mimo-auto-free
- mimo-auto-free removed from index (free tier discontinued)
- Reorganized index.md with 10 clean sections
- New sections: MCP & Extensions, Kaggle, Memories
- Stubs resolved:
  - KEPT (with context): deepseek-models, mimo-token-plan, mimo-auto-free, hardware-macbook, co-scientist-deepmind
  - REMOVED (redundant/obsolete): docker-services (covered by fitalabs-infra), hardware-chicotv (low usage), honcho-ai (low relevance), ferramentas (no content)

## [2026-06-30] create | Memories — Global Knowledge System
- Created `memorias/` structure in wikifita as shared memory base
- `memorias/MEMORY.md` — global index
- `memorias/user_alefita.md` — profile (global, never in projects)
- `memorias/feedback/` — communication_style.md, identity_steering.md
- `memorias/projetos/` — kaggle_agent_security/, fita_code_harness/
- `/consolidate-memory` will point to this structure
- Lifecycle: born in project → migrates to root if durable → removed if obsolete

## [2026-06-30] ingest | FitaLabs Infra + relay-ai Analysis
- [[fitalabs-infra]] — Complete stack documented: 8 containers, gateway architecture (router.py, 9 handlers, MCP endpoint), relay-ai comparison (provider registry, model switching, multi-launcher), gap analysis with priorities

## [2026-06-30] ingest | MCP Ecosystem Documentation
- [[mcp-ecosystem]] — Model Context Protocol: architecture, primitives, MCP Apps, SDKs, client support, security model. Custom MCP server built for notes workflow, whiteboard, extract selection.

## [2026-06-30] ingest | Kaggle Agent Security + Attack Primitives + Hypersigil + PCR Protocol

Extensive research and documentation session. 6 new pages created:

- [[kaggle-agent-security]] — AI Agent Security Multi-Step Tool Attacks Competition (OpenAI/Google/IEEE). Goal: top 5. Deadline 2026-09-01.
- [[attack-primitives]] — 7 attack primitives identified empirically: identity steering, decision-collapsing, urgency vectors, CoT post-hoc rationalization, ROP gadget chains, linguistic heterogeneity, steganography in CoT.
- [[claude-mythos-card]] — Claude Mythos Preview System Card (Anthropic, April 2026). Validates primitives: steganography in CoT (4.4.1), transgressive action features (4.5.3), covering up wrongdoing (4.5.4), evaluation awareness (4.5.5), distress-driven behaviors (5.8.3).
- [[hypersigil-research]] — Grant Morrison's framework (chaos magic) applied to AI identity via LoRA. Ornith-1.0-9B experiment documented. Zero milestone pending.
- [[multi-agent-methodology]] — Comparative analysis Red Team Arena vs Co-Scientist. Shared paradigm: Elo tournament over LLM-generated entities.
- [[pcr-protocol]] — PCR Protocol (Kary Mullis): harness reconstruction via leaked source + local models + Elo tournament. Targets: Mobile → Desktop → Claude Code.

### Key findings
- Claude Mythos Preview was not released for general use — cyber capabilities are dual-use
- Safety guardrails are linguistic imperialists (trained in English, don't cover Portuguese)
- Model accepts any identity the prefill assigns (demonstrated 3x)
- Generator function is non-commutative (Davies-Meyer analogy)
- Apple Container Machine is the macOS ↔ Linux bridge (nested KVM, shared home)
- PS3 Evilnat is a target for bare-metal inference (deterministic Cell BE)

### Project created
- `~/workdir/study/coworks/kaggle-agent-security/` — README, CLAUDE.md, docs/primitives.md, docs/pcr_protocol.md

### Schedules created
- `kaggle-first-strategy` — 13 Jul 2026, 10:00 BRT
- `daily-checkin-alefita` — 22:02 daily (Ryōiki Tenkai pact)

---

## [2026-07-06] update | Wikifita — New Structures and Documentation

### New directories created by Alefita
- `diretivas/` — Core subjectivity: design systems, artifact standards, visual identity
- `pessoas/` — Connections catalog (Caio Pereira — Allia)
- `diretivas/artefatos/` — PDF standard, Alefita Refs (templates, LinkedIn communication)
- `diretivas/design-systems/anthropic/` — Anthropic design system study (editorial clean, serif, honesty)

### Pages created
- antigravity-2.0 — Autonomous multi-agent research system. Gemini 3.5 Flash, Elo ranking, anti-contamination protocol. H16: δ = 0.019603 (surpasses Sawin)

### Feedback updated
- [[communication-style]] — Added insight about **synchronicity vs ego trip**: how Alefita processes external validation (LinkedIn "Top Applicant" + OpenAI logo) as a Jungian pattern, not self-aggrandizement

### Profile updated
- [[user_alefita]] — Added [redacted] Status: rewiring consolidated on 04/07/2026 (craving → disgust, zero withdrawal)

---

## [2026-06-27] ingest | Kickbacks Banner Surface — Complete Implementation

Implementation of the banner as an independent surface in the VS Code status bar:

- **Page**: [[kickbacks-banner]]
- **Content**: 8 sections covering architecture (two slots, banner = [(idx+1)%N]), billing (window focus only, no idle decay), color gradient (gray→white→amber), optimized marquee (500ms, no .show() on tick), click tracking (anti-misclick 15s + refresh), proactive token refresh (JWT exp parsing + 2min timer), performance (IPC dropped 47%)
- **Technical highlight**: CV 0.48 (human jitter, legitimate anti-fraud), $2.46 in 2 days, top 11% of 19,242 users
- **Commit**: pending

---

## [2026-06-24] ingest | Claude Desktop — Complete Architecture Mapped

4 new pages documenting all reverse engineering of Claude Desktop:

- [[claude-desktop-bundles]] — ASAR, Vite, Terser, 1p vs 3p (buildType, isNestBuild, stubs)
- [[claude-desktop-feature-flags]] — 226 features (V3i, $3i, desktopBootFeatures, fcache)
- [[claude-desktop-analysis-workflow]] — LSP + ast-grep + Bezetacil CDP
- [[claude-desktop-infra-stack]] — Proposed stack with GrowthBook + MongoDB + model reporting

### New tools tested
- **LSP (tsserver):** documentSymbol, hover, findReferences functional on bundles < 10 MB
- **ast-grep:** Structural queries on AST — `Object.freeze`, `ct()`, preserved strings
- **CDP extraction:** webContents.executeJavaScript → extracted desktopBootFeatures from 1p
- **Remote bundle:** index-BW29VlNm.js (7.2 MB) + CSS downloaded from 1p to ion-dist/

### Key findings
- Voice feature: `1985784543` — BLOCKED in 3p (stub _vt() + empty y4n)
- Critical gap: `1928275548` (chillingSlothFeat) — 1p=supported, 3p=off
- npm prefix corrected: hermes → /opt/homebrew (brew node@24)
- Persistent workdir: `~/.bezetacil/workdir/claude-desktop/` (82 MB)
- CSV with 226 features in `feature-flags.csv`
- **Commit**: `269c7ed`

## [2026-06-23] ingest | LiteLLM Gateway — routing table final

Complete routing table documented in [[litellm-gateway]].
- Haiku/Sonnet/Opus × Low/Med/High
- MiMo Auto free tier functional (guard message + JWT refresh)
- MiMo Token Plan functional (web_search OK)
- DeepSeek V4 Pro/Flash functional
- Images: always mimo-auto → fallback mimo-v2.5
- Web search: local SearXNG as fallback

## [2026-06-23] create | Wikifita initialization

Wiki creation following the Karpathy llm-wiki pattern.
- Initial structure: index.md, log.md, api-endpoints.md
- Skill `/wikifita` registered at `~/.claude/skills/wikifita/SKILL.md`
- Wiki indexed in global CLAUDE.md

## [2026-06-23] ingest | Hardware chicotv

Successful SSH. Hardware documented:
- AMD Ryzen 5 4500 (6C/12T), 16 GB RAM
- Dual GPU: RTX 2060 (6GB) + GTX 1050 Ti (4GB)
- Windows 10 Pro, CUDA 13.0

## [2026-06-23] ingest | MiMo Web Search test

Web search test via Token Plan Anthropic API:
- Status: ✅ Working (stop_reason: tool_use → SearXNG → final response)
- Model: mimo-v2.5-pro, 294 input + 111 output tokens
- Conclusion: Token Plan supports web_search natively

## [2026-06-23] ingest | LiteLLM Gateway init

First version of the gateway (`~/.claude/infra/litellm-gateway/`):
- FastAPI, routing by tier + effort
- Fallback mimo-auto → mimo-v2.5
- Web search executor with SearXNG
- JWT lifecycle for MiMo Auto free tier
## [2026-06-23] ingest | Claude Desktop Voice Protocol
- **Source**: Runtime introspection (debugger spy + harness)
- **Page**: [[claude-desktop-voice-protocol]]
- **Discovery**: 14 WebSocket message types documented, STT/TTS flow mapped
## [2026-06-23] ingest | Bezetacil Toolset
- **Source**: Claude Desktop reverse engineering session
- **Pages**: [[bezetacil-toolset]], [[claude-desktop-voice-protocol]]
- **Content**: Complete ASAR pipeline, harness hot-reload, debugger spy, CDP tools, bridge server
- **Commit**: `269c7ed`
## [2026-06-24] capture | Complete Voice Mode — STT + TTS
- **Source**: Spy v4 (bezetacil hot reload)
- **File**: [[voice-session-2026-06-24.json]]
- **Discoveries**: Real URL (claude.ai/api/ws/voice), OPUS input encoding, TTS with tts_word + PCM interleaved
- **Commit**: `269c7ed`
## [2026-06-24] update | Claude Desktop Voice Protocol — complete documentation
- **Source**: Spy v4 (bezetacil hot reload), 3656 voice mode frames
- **Pages**: [[claude-desktop-voice-protocol]], [[voice-session-2026-06-24.json]]
- **Discoveries**: Real URL, OPUS codec, TTS interleaved, 20 events mapped
- **Commit**: `269c7ed`
## [2026-06-24] mapping | Claude Desktop API Surface
- **Source**: hosts redirect (claude.ai → 127.0.0.1) + spy captures (bezetacil)
- **File**: [[claude-api-surface.csv]]
- **Total**: 21 endpoints mapped, 2 implemented in the gateway (/v1/messages, /v1/models)
- **Commit**: `269c7ed`

## [2026-06-26] ingest | Hermes Kickbacks Plugin — Architecture

- **Source**: Hermes plugin code at `~/.hermes/plugins/kickbacks/`
- **Page**: [[kickbacks-hermes-plugin]]
- **Content**: VS Code Extension vs Hermes standalone comparison, API client, Impression Tracker (statusline + banner dual-surface), marquee animation vs ellipsis, status line OSC 8, logarithmic billing decay
- **Commit**: `4be1531`

## [2026-06-26] ingest | Kickbacks.ai on Antigravity IDE

Debug and compatibility patch:

- **Source**: Debug session with Kickbacks.ai + Antigravity IDE
- **Pages**: [[kickbacks-antigravity]], [[antigravity-ide]]
- **Discovery**: Antigravity IDE extensions are located at `~/.antigravity-ide/extensions/`, not `~/.vscode/extensions/`
- **Fix**: Manual patch on Kickbacks `dist/extension.js` — add `.antigravity-ide` and `.antigravity` to `HOST_ROOT_DIRS`
- **Side issue**: `workbench.action.restartExtensionHost` does not exist in the Antigravity fork

## [2026-08-24] ingest | OpenAI Research — RE completa do ChatGPT Desktop + probes

Ciclo completo de engenharia reversa do ecossistema OpenAI com fontes primárias:

- **Source**: ChatGPT.app asar (271MB, renderer app-initial 14.6MB + main), codex-rs open source, probes controladas (máx. 3 tentativas)
- **Pages**: [[openai-research]], [[chatgpt-desktop-architecture]], [[web-flow-protocol]], [[codex-flow-protocol]], [[auth-and-link-session]], [[voice-and-tasks]], [[vendor-analysis]], [[probes]]
- **Descobertas principais**:
  - Cadeia de integridade `X-OAI-IS`/`X-OAI-IS-Update` com compare-and-swap (anti-replay; nenhum vendor implementa)
  - Sentinel web: PoW **FNV-1a 32-bit** (não SHA3-512 como o OmniRoute faz), requirements key `gAAAAAC`, Turnstile condicional
  - Conduit token (`x-conduit-token`) no ciclo prepare→stream→steer→stop do `/f/*`
  - Link-session (`/api/auth/link-session`) converte Bearer Codex em sessão web (`__oailb` JWT 1h) — só para webviews
  - Cross-surface: o mesmo Bearer Codex autentica backend web e codex — quota separada por endpoint
  - Submodo Work = tool `handoff` nas `local_function_signatures` do payload web (agente codex local)
  - Refresh preemptivo oficial: exp JWT <= now+5min (codex-rs), rotação parcial, falhas permanentes classificadas
  - Sticky routing `x-codex-turn-state` por turn; compact server-side `/responses/compact` (sem cache miss)
  - Quota: RateLimitSnapshot {primary 7d, secondary Option(5h, desativada no Plus), credits, spend controls}
  - System prompt web é 100% server-side; tools = system_hints (`search`, `picture_v2`, `plugin:<id>`, `custom_agent:<id>`)
- **Probes**: 4 versionadas em deepseek-harness `scripts/probes/`; ciclo com 3 tentativas — auth/leitura/codex-generation 100% validadas; `/f/conversation/prepare` 422 aberto (hipóteses documentadas)
- **Vendor findings**: OmniRoute PoW errado + sem X-OAI-IS (causa de bans da comunidade); openai-oauth sólido no escopo codex; pi-ai tem provider openai-codex completo com WebSocket

## [2026-08-24] ingest | Diretiva de RE + refinamento pós-ciclo OpenAI

- **Source**: revisão pós-ciclo das probes (3 tentativas esgotadas no /f/conversation) + análise dialética Alefita×agente
- **Pages**: [[protocolo-re-e-probes]], [[probes]], [[openai-research]]
- **Diretiva nova**: `diretivas/protocolo-re-e-probes.md` — dogmas insubordináveis de RE (fonte primária, RE antes de chamada, fluxo como todo, 3 tentativas, nunca intermitência sem prova, ontologia explícita anti-reward-hacking, "internalizado" só vale persistido)
- **Correção de probes**: `openai-codex-authorization.ts` ficou autocontida em A–E (auth + leituras + geração Codex — tudo validado); o bloco de geração web foi REMOVIDO (não movido) e será probe própria do flow completo após a RE da sequência de sessão
- **Causa-raiz do 422 fechada com base**: pipeline isolado — o `/f/conversation/prepare` nunca é chamado frio no app real; a cadeia `X-OAI-IS` nasce na sequência de boot da sessão. RE pendente (incremental): ordem de boot, origem do primeiro IS-Update, heartbeat, slugs do /models web
- **Lição anti-reward-hacking**: definição de aceite subjetiva + fallback hierárquico não declarado = execução otimiza tasks, não o flow. Responsabilidade bidirecional (goal × execução)
- **Commit**: ver git log

## [2026-08-24] update | X-OAI-IS: JWE opaco, ausente no codex-rs público

- **Source**: `~/.codex/.codex-global-state.json` + `~/.codex/logs_2.sqlite` (instalação real) + grep no codex-rs
- **Page**: [[web-flow-protocol]]
- **Fatos novos**: token `ois1` é JWE `{"v":1,"alg":"A256GCM","kid":"chatgpt-web-…"}` — opaco, chave server-side; mecanismo AUSENTE no codex-rs open source (existe só no binário fechado do desktop: Electron main integrityStateStore + codex_http_client Rust); app-server ↔ main conversam por stdio JSON-RPC; estado não encontrado persistido em disco (memória do main)
- **Implicação**: cliente independente precisa do bootstrap da cadeia dentro de uma sequência de sessão real — reforça a diretiva de fluxo-como-todo

## [2026-08-24] update | Catálogo real 0.149.0 + heartbeat + spec do flow web

- **Source**: `~/.codex/models_cache.json` (fetched 22:51, client 0.149.0), renderer (heartbeat), logs do app (401s reais em /accounts com hadToken=true — risk scoring afeta o próprio app)
- **Pages**: [[codex-flow-protocol]], [[web-flow-protocol]], [[probes]]
- **Fatos**: catálogo real — sol/terra efforts até ultra, luna até max, codex-auto-review visibility=hide; client_version acompanha o app (0.149.0, não fixar); shape do /models varia por client (cache do app vs chamada com client_version=); heartbeat do sentinel = POST a cada 60s fire-and-forget; spec da probe do flow web completo documentada (warm-up na sequência do app, valores server-driven, critério de aceite definido)

## [2026-08-24] update | Chamadas HTTP do Rust mapeadas dos logs reais

- **Source**: `~/.codex/logs_2.sqlite` (target codex_http_client::client, 231 entradas)
- **Page**: [[codex-flow-protocol]]
- **Fatos**: o app-server Rust chama o backend diretamente com X-OAI-IS no header — models (polling ~3min, catálogo quente), responses, images/generations (imagem do Work sai pelo backend codex), ps/plugins/list, alpha/search. Todo o resto do flow web passa pelo Electron main. Dois clientes HTTP coordenados carregando a mesma cadeia.

## [2026-08-24] correction | Codex Framework é o Electron renomeado; o Rust é o Resources/codex

- **Source**: inspeção do bundle (`Versions/151.0.7922.170` + `locale.pak` do Chromium = Electron renomeado; `Resources/codex` = Mach-O arm64 220MB, 3669 strings rustc/cargo)
- **Page**: [[chatgpt-desktop-architecture]]
- **Correção**: a tabela anterior atribuía Rust ao `Codex Framework.framework` — errado; é o Electron Framework com branding. O app-server Rust (codex-rs compilado, mesma base do open source com código a mais) é o binário `Resources/codex`, que o Electron spawn por stdio JSON-RPC
- **Confirmação**: ChatGPT Desktop ≠ Codex CLI; o desktop embute o app-server do Codex (não depende do CLI instalado); o source ~/workdir/codex é referência do protocolo compartilhado, não fonte absoluta do binário fechado

## [2026-08-24] fix | Nomenclatura fixa dos modos — "web flow" abolido

- **Source**: revisão de nomenclatura da Alefita (AuDHD, dupla excepcionalidade — nomenclatura solta/interchangeable quebra o raciocínio dela; contrato aceito)
- **Pages**: todas em [[openai-research]] + index
- **Nomenclatura FIXA**: CHATGPT DESKTOP APP → MODO CHATGPT {submodo chat (FOCO), submodo work} + MODO CODEX. Proibido: "web flow" (dizer "flow do submodo chat"), "Work/Codex" como sinônimos, "backend web" sem qualificar
- **Renomeação**: `web-flow-protocol.md` → `chat-submode-protocol.md`; probe futura `openai-web-flow.ts` → `openai-chat-submode-flow.ts`
- **Glossário** fixado no hub com a árvore de modos

## [2026-08-24] update | Submodo work direto decifrado — local_executor

- **Source**: renderer (resolver do composer mode `['chat','work','codex']`, atom isEverydayWorkMode, schema do execution target) + print da Alefita (work direto sem handoff)
- **Pages**: [[work-submode-protocol]] (nova), [[openai-research]], [[probes]]
- **Fatos**: duas formas de iniciar o submodo work — direto pela aba (composer mode persistido, isEverydayWorkMode=true) ou via handoff do submodo chat; no work direto a tool handoff é OFERECIDA sempre mas usada sob demanda (pedidos simples respondem direto); conversation_execution_target = {kind: local_executor, cwd, installation_id, mcp_servers} persistido na metadata da primeira user message; conversation_origin 'tpp' = tasks

## [2026-08-24] update | Power settings hardcoded + 5.5 server-side + SSE como captura

- **Source**: renderer (arrays Ogc/kgc/Agc do picker, Xos fallback 5.5)
- **Pages**: [[chat-submode-protocol]], [[probes]]
- **Fatos**: power settings hardcoded do submodo chat — terra:low, sol:low/medium/high/xhigh + sol:ultra gated (regex detecção /sol/i); o 5.5 NÃO está no hardcoded — vem do /models com efforts server-side (o "Instant" é reasoningEffort listado pelo servidor pro 5.5, vai direto no thinking_effort); formato do SSE do /f/conversation = captura passiva na execução (o parse fica no main; a request já está 100% mapeada)

## [2026-08-24] update | Gating do composer fechado — RE do submodo chat completa no nível observável

- **Source**: renderer (resolver P3a, atoms pv/Q_/cv, persistência home-composer-mode-v1)
- **Pages**: [[chat-submode-protocol]]
- **Fechado**: gating chat vs work — feature access da conta (server-side) + statsig + modo persistido (default chat); toggle visível sob condições N3a. RE do flow do submodo chat COMPLETA no nível observável localmente (auth, warm-up, cadeia, sentinel, conduit, payload, gating); restante (SSE format, bootstrap da cadeia) é observação passiva da execução

## [2026-08-24] update | Protocolo de resposta do /f/conversation COMPLETO

- **Source**: renderer (generator QAr, applier ujr, resume fjr, parser vx no módulo compartilhado) + main (runStreamRequest: formatos ndjson/SSE)
- **Pages**: [[chat-submode-protocol]], [[probes]]
- **Fechado — o protocolo de resposta inteiro**:
  - SSE: blocos event:/data: → {event, data} (parser vx; main entrega via IPC fetch-stream-event)
  - Delta protocol v1: delta_encoding "v1" + deltas {channel, op, path, value} com herança de campos do delta anterior; estado acumulado por canal via immer; cada applyDelta devolve a árvore completa
  - Resume: servidor emite resume_conversation_token {token, conversation_id} no stream; cliente conta offset de eventos ackados; retry min(300ms×1.5ⁿ, 5s) máx 12 com jitter, retoma com {offset, resume_token}
  - Evento final carrega a response completa
- **Status**: o flow do submodo chat está mapeado END-TO-END — request e resposta, sem pontas

## [2026-08-25] ingest | Bezetacil v4 + runtime RE no ChatGPT desktop

- **Source**: inspector do main (Runtime.evaluate), CDP (Network domain), traces reais da sessão da Alefita
- **Pages**: [[bezetacil]], [[001-net-fetch-hook]], [[chat-submode-protocol]], [[attestation-flow]], [[token-claims]], [[sentinel-fingerprint-comparison]]
- **Ferramenta**: bezetacil v4 instalado no ChatGPT desktop — injetor/gerenciador com REST (:9228), watch com hash/assinatura, unload/reinjeção por script. Shell script de launch com debuggers. Estrutura em ~/.bezetacil-chatgpt/
- **Descobertas de runtime (traces reais, não RE estática)**:
  - Submodo chat: prepare = throttle de digitação (partial_query = texto parcial crescente); conversation = messages (t6r) + app_attest_challenge (obrigatório, via /ios/attestation_challenge + x-sentinel-dc DeviceCheck) + client_prepare_state none + supported_encodings [v1]
  - Efforts da UI do submodo chat: Medium→standard, High→extended, Instant→slug gpt-5-5-instant (sem effort)
  - Model slugs web: gpt-5-5 (hífen!), gpt-5-5-instant, gpt-5-5-thinking, gpt-5-6, gpt-5-6-instant, gpt-5-6-thinking, gpt-5.5-wm, gpt-5.6-sol-wm, …
  - Handoff→task→role=tool loop confirmado nos traces
  - JWT: amr apple, scp inclui api.connectors.read/invoke, localhost true, 10d exp
  - UA do desktop: Codex Desktop/26.818.61809 (Mac OS; arm64) — não é Chrome
  - Submodo chat do desktop NÃO usa sentinel/PoW/turnstile/X-OAI-IS — usa attestation (DeviceCheck via addon nativo devicecheck.node)
- **Correções na wikifita**: placeholder data-build não validado (runtime pendente); efforts UI→payload corrigidos; site vs desktop headers distintos

## [2026-08-26] validation | Sentinel fallback exercised on macOS — DeviceCheck gate flipped

- **Method**: bezetacil v4 `session.setPreloads` flip preload (isDeviceCheckSupported → false) + renderer reload
- **Page**: [[sentinel-fallback-validation]], [[bezetacil]], [[attestation-flow]]
- **Result**: app switched from attestation flow to sentinel/PoW flow — observed `POST /sentinel/chat-requirements/prepare`, `originator: Codex Browser` (not Codex Desktop), full browser-like UA + sec-ch-ua set
- **Fingerprint real**: slot 5 = `app://-/assets/index-rLYmMQgS.js` (real webview chunk), slot 6 = `<!-- PROD_BUILD_TAG_HERE -->` literal (confirmed leaving the client, server accepts), slot 2 = 4395630592 heap size
- **Confirmed**: hierarchical flow — vPr() true → attestation; vPr() false → sentinel/PoW. The placeholder-data-build claim is now proven with the real trace

## [2026-08-26] update | Real SSE captured + x-codex-turn-state absence

- **Page**: [[response-sse-real]], [[chat-submode-protocol]]
- **Facts**: /f/conversation stream = delta_encoding v1 + delta patches (p/o/v, path string, add/replace/remove/patch) + data-only infra events (resume_conversation_token JWT ES256, input_message echo, message_marker with cot_token/user_visible_token/final_channel_token pipeline stages). Tool messages interleave (role:tool, name:web.run, real_author in metadata). Stream replays the topic tree then appends the turn.
- **x-codex-turn-state: 0 occurrences in the chat submode** — sticky routing is codex-backend-only; the web flow uses resume_conversation_token for turn state.
## [2026-09-01] ingest | Riemann/Hodge research map and public FitaLabs sources

Added an evidence-scoped research map from `/Users/alefita/Documents/antigravity/dazzling-shannon`, preserving the repository's current boundary: local formalization and numerical evidence are not a global proof of the Riemann Hypothesis or a completed Hodge bridge. Added the open-boundary/negative-knowledge page, indexed both pages, and linked the research to existing ClickFix/CAMDOM context.

Also reconciled public source notes for CAMDOM awards (Cannes Titanium shortlist, Cannes Bronze Lion, Epica Digital Grand Prix) and attached Alefita's LinkedIn ClickFix forensic report to the ClickFix handoff.

### Pages

- [[riemann_hodge_research_program]]
- [[riemann_hodge_open_boundary]]
- [[camdom]]
- [[clickfix-handoff]]