From dc324db374ddf79a2dc07f4823e0334c3ba6a393 Mon Sep 17 00:00:00 2001 From: Claude Date: Fri, 10 Jul 2026 16:18:30 +0000 Subject: [PATCH 1/4] board: E-NOBODY-WAITS-1 banked + #674 post-merge arc entry (doc-only) Operator ruling from the kanban-step-vs-ack-pump audit: no messages, no actors, anywhere - ractor is ONLY the compile-time ownership guarantee. The canonical loop is message-free (ack_and_propose proposal -> MailboxSoaOwner::try_advance_phase(&mut); &mut IS the serialization - the codex #578 stale-phase race and #579 exec-overlay both dissolve under exclusive ownership). The Rubicon model is intact and owner-side (cast records intent AHEAD of the ack - crossing at intent formation). The supervisor's KanbanMsg/ractor::call! drivers are regraded as documented redundancy - operator disposition LEAVE AS IS (TD-MESSAGE-RESIDUE; no retirement queued; drift signal = NEW code reaching for messages). Prime invariant, the operator's closing emphasis: NOBODY WAITS FOR ANYTHING OR ANY SCHEDULING - writes fire-and-forget, acks propose, updates reprioritize never gate, absorbing views rest never deadlock. Also: PR_ARC_INVENTORY #674 entry prepended (merge cd5178e - the deferred post-merge hygiene); AGENT_LOG entry. No code changes. Co-Authored-By: Claude --- .claude/board/AGENT_LOG.md | 7 +++++++ .claude/board/EPIPHANIES.md | 10 ++++++++++ .claude/board/PR_ARC_INVENTORY.md | 16 ++++++++++++++++ .claude/board/TECH_DEBT.md | 17 +++++++++++++++++ 4 files changed, 50 insertions(+) diff --git a/.claude/board/AGENT_LOG.md b/.claude/board/AGENT_LOG.md index 0003f199..9aa6646c 100644 --- a/.claude/board/AGENT_LOG.md +++ b/.claude/board/AGENT_LOG.md @@ -1,3 +1,10 @@ +## 2026-07-10 — fable-674-postmerge — #674 MERGED; E-NOBODY-WAITS-1 banked (doc-only, operator: "leave it as is and just document") + +- **#674 merged** (`cd5178e`); branch restarted from main; PR_ARC_INVENTORY #674 entry prepended (the deferred post-merge hygiene, now done). +- **Operator audit resolved:** "kanban step vs ack pump — duplicate route? Rubicon ignored?" → grounded answer: Rubicon intact and owner-side (KanbanColumn IS the 4-phase model; cast records intent AHEAD of ack — crossing at intent formation); the ack-pump is the canonical MESSAGE-FREE route (`ack_and_propose` proposal → `try_advance_phase(&mut)`; `&mut` is the serialization — codex #578's race and #579's overlay both dissolve); rs-graph-llm has zero ack code (storage-side concern only). The redundancy runs the OTHER way: supervisor `KanbanMsg`/`ractor::call!` drivers are a second, message-based model. +- **Ruling banked:** E-NOBODY-WAITS-1 (no messages, no actors anywhere; ractor = compile-time ownership guarantee only; **prime invariant: nobody waits for anything or any scheduling**). Disposition per operator: LEAVE AS IS — TD-MESSAGE-RESIDUE documents the residue, no retirement queued; drift signal = NEW code reaching for KanbanMsg/call!. +- Doc-only commit: EPIPHANIES + TECH_DEBT + PR_ARC + this entry. No code touched. + ## 2026-07-10 — fable-dmts6-probe — D-MTS-6 measured GREEN + PR #674 opened + codex P2 ack-dedup fix - **PR #674 opened** (operator ask): the full 14-commit branch arc; session subscribed to PR activity; hourly self check-in armed. diff --git a/.claude/board/EPIPHANIES.md b/.claude/board/EPIPHANIES.md index 89c209e6..610b6943 100644 --- a/.claude/board/EPIPHANIES.md +++ b/.claude/board/EPIPHANIES.md @@ -3,6 +3,16 @@ **The probe** (zero-dep, D-MTS-5 machinery: seed `0x9E3779B97F4A7C15`, M=4096, COMMA_STRIDE=2395, L=12 levels; baseline = the full CausalEdge64's 8+8-bit truth payload): each level stores only k bits of each truth value with the comma phase acting as **deterministic subtractive dither** (generated from the address, never stored); read-time de-dithers and averages the L witnesses. Three awareness proxies vs the baseline over 8192 NARS revision/deduction pairs + a 64-step revision chain: revision-outcome |ΔE|, MetaWord-analog surprise-bit agreement, F-descent trajectory ρ. **Measured (ALL GATES PASS):** the comma lane's knee is **k\*=1** — one stored bit per level (2 explicit truth bits/edge vs the baseline 16) already gives mean|ΔE| 0.0084, surprise agreement 0.9688, descent ρ 0.9792. The aligned control (collapsed pyramid) needs **k\*=4**. Mechanism made visible: the comma walk is a **low-discrepancy per-edge lattice**, so de-dithered averaging behaves like stratified sampling — comma k=1 RMSE 0.0244 vs aligned 0.2503 (10×), matching aligned between k=3 and k=4: the lattice buys **≈3.4 effective bits ≈ log₂(12)=3.58**. Replay bit-identical across passes and level orders. **Honest chronicle:** run #1 FAILED its own G1 — the gate was mis-registered (demanded EXACT surprise agreement at k=8, but the comma lane is a dithered reconstruction, never an identity path); the fix was a diagnosis measurement, not a tune: G1′ additionally requires every disagreeing pair's margin-to-threshold < 0.005, and the measured max margin is **1.7e-5** — every flip is threshold-boundary noise, proven. **Fences:** proxies ≠ the full ShaderDriver loop — any real CausalEdge64 shrink is gated on **D-MTS-6b** (driver-integrated fixture); the economy claim is marginal-cost (the L per-level slots are the pyramid's OWN envelope storage — the probe measures what they jointly RECOVER, not free storage). +## 2026-07-10 — E-NOBODY-WAITS-1 — no messages, no actors, anywhere: ractor is ONLY the compile-time ownership guarantee; the prime invariant is that NOBODY waits for anything or any scheduling +**Status:** RULING (operator, 2026-07-10, post-#674-merge audit of "kanban step vs ack pump"; disposition: LEAVE AS IS AND DOCUMENT — no retirement work opened) + +The operator's audit question — did the ack-pump duplicate the existing BatchWriter→kanban-step route, or ignore the Rubicon model? — resolved into a foundational clarification: **we don't use messages at all, nor actors; ractor exists solely as the compile-time ownership guarantee** (the delegation declaration — `KanbanActor`-over-real-`MailboxSoA` names WHO owns, never a message path). Consequences, all verified in code: + +1. **The Rubicon model is intact and owner-side.** `contract::kanban::KanbanColumn` IS the 4-phase Rubicon (Planning→CognitiveWork = the −550ms Σ-commit crossing; Planning→Prune = the pre-Rubicon Libet veto; Evaluation 3-way; Plan→Planning re-deliberation). The crossing is the owner's, at intent formation — the cast records intent AHEAD of the ack, Rubicon-conformant. +2. **The message-free loop is the canonical route:** `BatchWriter::ack_and_propose` (first-transition-dedup'd, #674) → `Option` proposal → `MailboxSoaOwner::try_advance_phase(&mut self, to)` — plain method calls end to end. **`&mut` IS the serialization**: the stale-phase race codex #578 guarded against, and the atomicity ceremony built for it, only exist in a message world; under exclusive ownership the race is a compile error — the ractor ruling doing its job. The #579 exec-overlay concern also dissolves in the direct path (the proposal itself is applied; `exec` travels with it). The ack is storage-side postactional evidence of cycle N seeding cycle N+1's deliberation — it proposes, the owner disposes, and rs-graph-llm contains no ack code at all (consumer transitions are envelope KanbanMoves on the same Rubicon DAG). +3. **Any message path in the tree is redundancy.** Named residue: `lance-graph-supervisor`'s `KanbanMsg::{Tick,Advance}`, `ractor::call!` RPC, `drive_version_tick`, `drive_scheduled_tick` (consumers: supervisor tests + onebrc-probe lane E — which, fittingly, measured the actor-model tax). **Operator disposition: left as is, documented** (TECH_DEBT TD-MESSAGE-RESIDUE); the drift signal is NEW code reaching for `KanbanMsg`/`call!` instead of the `&mut` owner surface. +4. **THE PRIME INVARIANT (the operator's closing emphasis): nobody is waiting for anything or any scheduling.** Writes are fire-and-forget ("melden macht frei"); acks propose; updates reprioritize, never gate; absorbing views rest the loop, never deadlock it; the thinker never blocks on I/O, on a scheduler, or on another mailbox. Every future design choice on this surface is judged against this sentence first. + ## 2026-07-10 — E-FIVE-STYLE-TABLES-1 — the M9 dedup found FIVE mutually divergent hand-rolled style tables; duplication had already produced live semantic drift, and the weakest gate would have greenlit silent corruption **Status:** FINDING (measured during D-TSC-1 implementation, the first 5+3 council run; all five replaced by `StyleFamily::default_runbook()`/`ThinkingStyle::family()`, shipped with 1549 tests green) diff --git a/.claude/board/PR_ARC_INVENTORY.md b/.claude/board/PR_ARC_INVENTORY.md index a6a2d7b3..b7906667 100644 --- a/.claude/board/PR_ARC_INVENTORY.md +++ b/.claude/board/PR_ARC_INVENTORY.md @@ -35,6 +35,22 @@ --- +## #674 lance-graph: V3 W2–W6 continuation + comma quorum/awareness measured + 5+3 council + StyleFamily dedup (D-TSC-1) + +**Status:** MERGED 2026-07-10 (merge commit `cd5178e`), branch `claude/review-claude-board-files-nhqgx1`, 16 commits. Companion: rs-graph-llm `8ef18b9` (same branch name there, not yet PR'd) — `run_cycle` explicit `on_behalf` (owner ≠ classid). + +**Added:** `contract::step_mask::StepMask` (W3a); `planner::elevation::cycle::CycleBudget` (W2d/M12); `MailboxSoA::cast_on_behalf` (W4a) + standalone `with-planner` E0432 fix; `contract examples/adoption_scan` (W6a); `BatchWriter::ack_and_propose` (D-ORG-1, first-transition-dedup'd per codex P2); `contract::style_family::StyleFamily` + `default_runbook()`/`ThinkingStyle::family()` (D-TSC-1/M9); `PlannerStyleExt`/`EngineStyleExt`; `superposition::DetectedStyle` rename; thinking-engine gains the lance-graph-contract dep; probes `perturbation-sim/examples/{comma_quorum,comma_awareness}.rs` (D-MTS-5/6); `.claude/agents/5plus3-council.md` + `/5plus3` skill; `.claude/v3/FUTURE-DESIGN.md`. + +**Locked:** the seven 2026-07-10 rulings (E-ACK-IS-THE-KANBAN-TRIGGER-1, E-ORCHESTRATION-ORGANS-1, E-THINKING-STYLES-ARE-CLASSES-1, E-MARKOV-TEMPORAL-STREAM-1, E-COMMA-QUORUM-1/E-COMMA-REPLAY-1, E-STYLE-FAMILY-VS-RUNBOOK-1, E-THINKING-TENANTS-V3-1); measured findings E-COMMA-QUORUM-MEASURED-1 (N_eff 11.00/12; boundary condition N_eff = min(L, spectral participation)), E-COMMA-AWARENESS-MEASURED-1 (k*=1 vs aligned k*=4; ≈log2(12) effective bits), E-FIVE-STYLE-TABLES-1 (five divergent style tables — the M9 payoff made concrete); M9 RESOLVED with G1 = 1 enum + 3 deprecated aliases; the 5+3 anti-mush sequencing (spec-first, consolidate-before-review); style behavior changes documented + G7 literal-pinned (planner arms 9/10/11, driver 8/9/10, parse 8/9/10, engine 36→12 ranges). + +**Deferred:** D-MTS-6b (driver-integrated awareness fixture — gates ANY real CausalEdge64 shrink); D-MTS-1..3 (gate ALL VSA-path removal); D-TTV-1 (thinking tenants → V3); D-TSC-2/3 (batched OGAR mint, never solo); D-TSC-4 (W6c/0x1000 coexistence — ESCALATED, operator ruling needed); W2a BoardAggregates (batched-mint gate); TD-STYLE-TABLE-RESIDUE (ndarray pair / wip imports / p64-bridge ordinal probe). + +**Docs:** spec `dtsc1-thinkingstyle-dedup-spec-v1.md` (v1→v2→v3, full council fix ledgers); plan `temporal-markov-and-style-classes-v1.md`; FUTURE-DESIGN landing zone; TYPE_DUPLICATION_MAP/COMPONENT-MAP/MODULE-TABLE addenda; .grok thinking_styles superseded banner. + +**Post-merge ruling (2026-07-10, same day):** E-NOBODY-WAITS-1 — no messages/no actors anywhere (ractor = compile-time ownership guarantee only); the ack-pump is the canonical message-free route; supervisor KanbanMsg drivers regraded documented redundancy (TD-MESSAGE-RESIDUE, leave-as-is); prime invariant: nobody waits for anything or any scheduling. + +**Confidence (2026-07-10):** HIGH — all 6 CI checks green at merge; 1549 lib tests across the four touched crates; both probes ALL GATES PASS with honest run chronicles. + ## #632 lance-graph: Cross-session intake — RouteBucketTyped (C6) merged, emission_scan minted, OCR codebook mirror, GraphRAG-rs inventory + operator rulings **Status:** MERGED 2026-07-02 (merge commit `df367471`), branch `claude/v3-substrate-migration-review-o0yoxv`. Companion: OGAR #148 (merge `75d955b`) — the cross-repo fuses + serialized allocation batch; merged FIRST per the lockstep note; this post-merge commit carries the ogar-vocab lock bumps that cleared COUNT_FUSE (fired 65 vs 68 in the in-between state — first live demonstration of the fuse mechanism, then green 68 == 68, lance-graph-ogar 81 tests). diff --git a/.claude/board/TECH_DEBT.md b/.claude/board/TECH_DEBT.md index 8f1e0ac6..abbc01eb 100644 --- a/.claude/board/TECH_DEBT.md +++ b/.claude/board/TECH_DEBT.md @@ -1,5 +1,22 @@ # Technical Debt Log — Open + Paid (double-entry, append-only) +## TD-MESSAGE-RESIDUE (2026-07-10, operator-ruled LEAVE-AS-IS) + +Per E-NOBODY-WAITS-1: no messages, no actors anywhere — ractor is only +the compile-time ownership guarantee. The message-based arm in +`lance-graph-supervisor` (`KanbanMsg::{Tick,Advance}`, `ractor::call!`, +`drive_version_tick`, `drive_scheduled_tick`; consumers: supervisor +tests + onebrc-probe lane E) is REDUNDANCY relative to the canonical +message-free loop (`BatchWriter::ack_and_propose` proposal → +`MailboxSoaOwner::try_advance_phase(&mut)` — `&mut` is the +serialization). **Disposition: left as is, documented — NO retirement +work queued** (operator, 2026-07-10). Not a bug: the code is correct in +its own model; it is a second model. Drift signal: NEW code reaching +for `KanbanMsg`/`ractor::call!` instead of the `&mut` owner surface — +that, not the existing residue, is what a reviewer blocks. The prime +invariant every change on this surface is judged against: **nobody +waits for anything or any scheduling.** + ## TD-STYLE-TABLE-RESIDUE (2026-07-10, D-TSC-1 follow-ups) Three residues from the M9 ThinkingStyle dedup (all OUT of D-TSC-1 scope, From 13188ad0cec2b6f863b6259b7479cff37ebee3a4 Mon Sep 17 00:00:00 2001 From: Claude Date: Fri, 10 Jul 2026 19:55:36 +0000 Subject: [PATCH 2/4] v3: VISION.md draft + ancestry census + arm/tenant design input (filigree in flight) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - .claude/v3/VISION.md — the AGI-aspiring canon re-grounded (draft: two filigree reviews commissioned; dilution/collapse verdicts are in (1x FIX-P1: organ table must add free_energy + global_context rows; 4x FIX-P2), overclaim review pending; fixes land in the follow-up commit per the version-ledger pattern). - MODULE-TABLE addendum (2): ancestry-pipeline census — 73 files across thinking-engine (49) / p64-bridge (1) / cognitive-shader-driver (23), with gem-status column. Notable finds queued for the residue ledger: style tables beyond the D-TSC-1 sweep (auto_style u8 consts, UNIFIED_STYLES params, p64-bridge StyleParams), GestaltState defined twice, lens-module triplet, wire.rs largest driver file while LAB-ONLY. - plan temporal-markov-and-style-classes-v1 addendum (third wave, operator): AriGraph context V3-TENANT-SHAPED for Markov context building while streaming text; ingest legs = lance-graph-arm-discovery (Aerial+ transcode, palette256 oracle) + DeepNSM (4096 COCA = CAM codebook); the D-MTS-1 spec must weigh CAM-PQ 6x8=48-bit path codes (address side) vs 6x palette256:palette256 = 12 B = ONE V3 tenant (value side); likely resolution: key addresses, tenant carries. - AGENT_LOG census-worker entry. Co-Authored-By: Claude --- .claude/board/AGENT_LOG.md | 6 + .../temporal-markov-and-style-classes-v1.md | 25 ++ .claude/v3/MODULE-TABLE.md | 144 ++++++++++++ .claude/v3/VISION.md | 217 ++++++++++++++++++ 4 files changed, 392 insertions(+) create mode 100644 .claude/v3/VISION.md diff --git a/.claude/board/AGENT_LOG.md b/.claude/board/AGENT_LOG.md index 9aa6646c..80824b74 100644 --- a/.claude/board/AGENT_LOG.md +++ b/.claude/board/AGENT_LOG.md @@ -1,3 +1,9 @@ +## 2026-07-10 — census-worker-ancestry-pipeline — V3 census extended to thinking-engine / p64-bridge / cognitive-shader-driver (doc-only) + +- **Census worker (Sonnet, doc-only, edit-only fleet rule honored):** appended `.claude/v3/MODULE-TABLE.md` ADDENDUM covering 49 thinking-engine files + p64-bridge/lib.rs + 23 cognitive-shader-driver src files (bare `crates/p64` does not exist as a workspace member — only `p64-bridge`). Added a `gem-status` column (WIRED-HOT-PATH / UNWIRED-GEM / CALIBRATION-ONLY / LAB-ONLY / RESIDUE) per FUTURE-DESIGN.md's wiring-queue framing. +- **Notable finds:** style-table proliferation is WORSE than FUTURE-DESIGN/D-TSC-1 tracked — `superposition.rs::ThinkingStyle = DetectedStyle` collides in name (not value) with `cognitive_stack.rs::ThinkingStyle = StyleFamily` in the SAME crate; `auto_style.rs` (bare consts) + `engine_bridge.rs::UNIFIED_STYLES` + p64-bridge's `StyleParams` are 3 MORE independent 12-slot tables beyond what D-TSC-1's fleet-run addressed — worth a follow-up grep to confirm `UNIFIED_STYLES`'s definition site (not just callers) was routed. Also: `GestaltState` independently defined in both `awareness_dto.rs` and `world_model.rs`; 3 near-identical lens modules (jina/bge_m3/reranker_lens); `signed_domino.rs`/`branching.rs` extend the M8 near-duplicate-engine count beyond the 3 FUTURE-DESIGN names. `splat_ops.rs` is RESIDUE (self-declared deprecated, scheduled sprint-15+ removal, do not touch per §1 rule 8). +- **No code changed.** Working tree has the MODULE-TABLE.md addendum only; not committed. + ## 2026-07-10 — fable-674-postmerge — #674 MERGED; E-NOBODY-WAITS-1 banked (doc-only, operator: "leave it as is and just document") - **#674 merged** (`cd5178e`); branch restarted from main; PR_ARC_INVENTORY #674 entry prepended (the deferred post-merge hygiene, now done). diff --git a/.claude/plans/temporal-markov-and-style-classes-v1.md b/.claude/plans/temporal-markov-and-style-classes-v1.md index b84ce25d..e322fa0f 100644 --- a/.claude/plans/temporal-markov-and-style-classes-v1.md +++ b/.claude/plans/temporal-markov-and-style-classes-v1.md @@ -59,6 +59,31 @@ never stored); per-level content = palette-quantized magnitude envelope only. > E-COMMA-AWARENESS-MEASURED-1. > Execution spec for D-TSC-1: `.claude/plans/dtsc1-thinkingstyle-dedup-spec-v1.md`. +> **Addendum (operator, 2026-07-10, third wave — D-MTS-1 context-building +> design input):** the **AriGraph context, V3-TENANT-SHAPED, is most +> probably essential for Markov-chain context building while streaming +> text.** The ingest legs feeding it: **`lance-graph-arm-discovery`** +> (the Aerial+ transcode — float-free association-rule discovery via the +> palette256 integer distance oracle, ρ=0.9973, → `{s,p,o,f,c}` ndjson → +> SPO loader) and **DeepNSM** (text→SPO via the 6-state PoS FSM; the +> **64×64 = 4096 COCA vocabulary IS the CAM index codebook**, 4096² u8 +> LUT). The representational comparison the D-MTS-1 spec MUST weigh for +> the streamed context items: +> +> - **CAM-PQ `6×8 = 48-bit` codes** — 6 subspaces × 1 byte = 6 bytes = +> exactly the OGAR GUID path (HEEL+HIP+TWIG): the ADDRESS-side coding. +> - **`6× palette256:palette256` = 12 bytes = ONE V3 tenant** — 6 +> subspaces × an X:Y byte PAIR each (the 256×256 centroid-tile +> reading, le-contract L4 carve): the VALUE-tenant-side coding — +> exactly one content-blind facet payload. +> +> The likely resolution (to be probed, not assumed): key addresses, +> tenant carries — the 48-bit path codes route/dedupe the streamed SPO +> items, while the palette² tenant is the projection surface the +> perturbation shader attends per cycle. This joins D-TTV-1 (the +> AriGraph-context lane is a thinking tenant) and instantiates the L4 +> carrier named in E-MARKOV-TEMPORAL-STREAM-1. + ## Track B — Thinking styles as classes (domain:appid:classview) Ruling: styles move entirely under the classid umbrella. A style class diff --git a/.claude/v3/MODULE-TABLE.md b/.claude/v3/MODULE-TABLE.md index caf17afb..71422d76 100644 --- a/.claude/v3/MODULE-TABLE.md +++ b/.claude/v3/MODULE-TABLE.md @@ -344,3 +344,147 @@ yet censused); planner style.rs = re-export + `PlannerStyleExt`; `parse_style_name` routes the 12 family names through `StyleFamily::default_runbook()` (3 arms changed — the fourth divergent table); `nars_engine.rs` gained the runbook-keyed `style_vector_for`. + +## ADDENDUM 2026-07-10 (2) — ancestry-pipeline census: thinking-engine / p64 / p64-bridge / cognitive-shader-driver + +> Sonnet census worker. Depth: header doc-comment + `pub` surface grep per +> file (no deep body reads). Same columns as MODULE-TABLE.md rollup, plus +> **gem-status**: WIRED-HOT-PATH / UNWIRED-GEM / CALIBRATION-ONLY / +> LAB-ONLY / RESIDUE. Cross-ref `FUTURE-DESIGN.md` § "the migration arc" for +> the prior gems queue — this census confirms those and adds several not +> previously named. +> +> **Note on scope:** a bare `crates/p64` does not exist in this workspace — +> only `crates/p64-bridge` is a workspace member (`Cargo.toml` line 33). p64 +> itself is presumably the sibling `ndarray`-hosted crate consumed by name +> (`p64_bridge::CognitiveShader` docstrings reference "p64 (ndarray)" as the +> topology host) — not censused here since no local crate dir exists to walk. + +### thinking-engine (49 files, `crates/thinking-engine/src/*.rs`) + +| File | Consumes | Emits (key types) | LE contract | Function | Tech debt / duplication | gem-status | +|---|---|---|---|---|---|---| +| `lib.rs` | all submodules | module wiring | none | crate root, doc: "16M RISC Thought Engine" | none | WIRED-HOT-PATH | +| `auto_detect.rs` | config.json / GGUF metadata | `Architecture`, `DetectedModel` | none | routes model config → codebook/tokenizer/lens choice | none | CALIBRATION-ONLY | +| `awareness_dto.rs` | superposition/qualia results | `ResonanceDto`, `QualiaDto`, `MomentDto`, `GestaltState` | none | structured cognitive-state DTOs (perspectival Resonance — KEEPS its name per CLAUDE.md D-PERT-1 split) | none | WIRED-HOT-PATH (DTO ladder) | +| `bf16_engine.rs` | `bgz_tensor::stacked_n::ClamCodebook` | `BF16ThinkingEngine` | BF16 u16 table, lossless bit-shift decode | one-MatVec-per-cycle engine at BF16 precision | none | WIRED-HOT-PATH | +| `bge_m3_lens.rs` | baked 256×256 BF16 table (XLM-R/BGE-M3) | `bge_m3_lookup/_many`, `bge_m3_distance`, `bge_m3_engine` | none | one of 3 near-identical per-model lens modules (see jina_lens/reranker_lens) | duplication: shares N_CENTROIDS/VOCAB_SIZE/lookup/distance/engine shape verbatim with `jina_lens.rs` + `reranker_lens.rs` — not in FUTURE-DESIGN's gem list, candidate for a `Lens` generic | WIRED-HOT-PATH | +| `branching.rs` | none (self-contained) | `BranchingEngine`, `BranchResult`, L1/L2/L3 consts | none | 4×4 branching parallel-vector cascade (NOT filtering, spawns not survives) — a 4th cascade shape alongside domino/signed_domino/layered | not named in FUTURE-DESIGN's engine-collapse list; likely 5th near-duplicate cascade candidate | UNWIRED-GEM (not referenced by FUTURE-DESIGN's M8 list) | +| `bridge.rs` | `highheelbgz::SpiralAddress` | `spiral_to_table_index`, `build_spiral_distance_table`, `hydrate_and_cosine` | none | maps SpiralAddress → flat NxN table index, hydrated cosine refinement | none | CALIBRATION-ONLY | +| `bridge_gate.rs` | none (trait-only) | `CognitiveBridgeGate` trait, `PassthroughGate`, `DenyAllGate`, `CognitiveOpKind` | none | cross-tenant authorization injection point; production impl lives in lance-graph-callcenter | doc explicitly says "lives here so thinking-engine remains lower-level" — RBAC-adjacent, not cited in FUTURE-DESIGN | UNWIRED-GEM (auth seam, not yet in migration arc docs) | +| `builder.rs` | `Lens`, `TableType`, `Pooling` | `ThinkingEngineBuilder`(implied), `BuiltEngine`, `Temperature`, `ThinkingPreset`, `CommitSink` | none | fluent construction API composing lens/table-type/pooling | none | WIRED-HOT-PATH | +| `centroid_labels.rs` | none (data-only) | `JINA_CENTROID_LABELS: [&str; 256]` | none | auto-generated English labels for 256 CLAM centroids | none | CALIBRATION-ONLY | +| `codebook_index.rs` | none | `CodebookIndex` | none | token_id → centroid-row index, save/load | none | WIRED-HOT-PATH | +| `cognitive_stack.rs` | `StyleFamily` (contract, post D-TSC-1) | `LayerId`(L1-L10), `ThinkingStyle = StyleFamily` alias, `EngineStyleExt`, `StyleParams`, `GateState`, `RungLevel`, `MetaCognition` | none | 10-layer cognitive stack + 12-style dispatch, migrated from ladybug-rs | `ThinkingStyle = StyleFamily` type alias — post-council canonical routing | WIRED-HOT-PATH | +| `cognitive_trace.rs` | superposition field | `SpoTriple`, `CognitiveTrace`, `LensTrace` | none | full provenance chain text→thought, appendable to KG | none | UNWIRED-GEM (provenance/debug tooling, not cited elsewhere) | +| `composite_engine.rs` | `BuiltEngine` (builder.rs) | `CompositeEngine`, `CompositeResult` | none | multi-lens composition via energy superposition | one of 4 near-duplicate engines (FUTURE-DESIGN M8) | UNWIRED-GEM (named in FUTURE-DESIGN) | +| `contract_bridge.rs` | `lance-graph-contract` 36-style taxonomy | `contract_style_to_engine`, `cluster_to_archetype`, `CascadeConfig`, `FastBusDto` | `FastBusDto::SIZE` (mem::size_of) | bridges contract's 36 ThinkingStyles → engine's 12 + cascade params | `contract_style_to_engine` was the FIFTH divergent style table found during D-TSC-1 impl (per AGENT_LOG); now routes via `StyleFamily::family()` | WIRED-HOT-PATH (post D-TSC-1 fix) | +| `contrastive_learner.rs` | static distance table | `ContrastiveLearner`, `LearnerStats` | none | online EMA table update from real cosine pairs (Stage 2A) | none | UNWIRED-GEM (online-learning path, no driver caller found in this census pass) | +| `cronbach.rs` | multi-lens distance vectors | `cronbach_alpha`, `CronbachResult`, `QuorumLevel` | none | Cronbach α multi-lens internal-consistency calibration | none | CALIBRATION-ONLY (confirmed FUTURE-DESIGN gem) | +| `domino.rs` | `ThinkingEngine` | `DominoCascade`, `CascadeAtom`, `Transition`, `DissonanceProfile`, 8 `CH_*` channel consts | none | top-K NARS-context cascade (unsigned); 1 of 4 near-dup engines | FUTURE-DESIGN M8 named | UNWIRED-GEM | +| `dto.rs` | engine outputs | `SourceType`, `ThinkingScale`, `StreamDto`, `PerturbationDto`, `ResonanceDto = PerturbationDto` (alias), `BusDto`, `ThoughtStruct`, `ThoughtIndex` | none | the canonical Φ/Ψ/B/Γ DTO ladder (CLAUDE.md "Thinking is a struct") | note: `ResonanceDto` alias here is the PERTURBATION one (Ψ), distinct from `awareness_dto.rs::ResonanceDto` (perspectival) — the exact D-PERT-1 split named in workspace CLAUDE.md | WIRED-HOT-PATH | +| `dual_engine.rs` | two `BuiltEngine`s | `DualEngine`, `DualResult` | none | side-by-side u8-CDF vs BF16 disagreement measurement | 1 of 4 near-dup engines (FUTURE-DESIGN M8) | UNWIRED-GEM | +| `engine.rs` | codebook centroids | `ThinkingEngine`, `CODEBOOK_SIZE`=4096, `TABLE_SIZE`, `quantize_energy_f32_to_i8` | none | the canonical u8 MatVec engine, incl. `cycle_vnni`/`cycle_auto` SIMD dispatch tiers | none | WIRED-HOT-PATH | +| `f32_engine.rs` | f32 codebook centroids | `F32ThinkingEngine`, `SparseBranchGraph` | none | full-precision (r=0.9999) signed MatVec, no quantization loss | none | WIRED-HOT-PATH | +| `ghosts.rs` | past thought peaks | `GhostType`, `Ghost`, `GhostField` | none | Friston free-energy persistent priors biasing future cascades | FUTURE-DESIGN gem: persona/archetype Layer-2 role-catalogue candidate | UNWIRED-GEM | +| `ground_truth.rs` | candle forward pass (Jina v5/BERT) | `calibration::{GroundTruthEmbedding, CalibrationCorpus, GroundTruthSource}`, `spearman_rank_correlation` | none | RAW ground-truth embeddings for calibrating baked tables | none | CALIBRATION-ONLY (confirmed FUTURE-DESIGN gem) | +| `inference_backend.rs` | none (trait registry) | `InferenceBackend` trait, `EncodedState`, `BackendGrade`, 7 backend structs (Passthrough/RaBitQ/Spiral/I8Hybrid/HhtlF32/Cascade/Base17Signature), `all_backends()` | none | runtime-switchable dispatch across ALL research codec paths simultaneously; explicit "nothing is killed, deprecation is data-driven" doctrine | none | UNWIRED-GEM (R&D bench registry, not cited in FUTURE-DESIGN) | +| `jina_lens.rs` | baked 256×256 table (Jina v3/XLM-R) | `jina_lookup`, `jina_distance`, `jina_engine`, `jina_think` | none | RESEARCH-ONLY legacy lens (doc explicitly says do not use for new production wiring; Jina v5 is ground truth) | doc self-flags as superseded; same shape as bge_m3_lens/reranker_lens | RESIDUE (self-declared legacy) | +| `l4.rs` | centroid vectors | `L4Experience`, `ACCUM_LEN`=16384, `BIN_BYTES`=2048, `xor_bind`, `recognize`, `bias_sensor` | XOR-bound i8 accumulator, save/load byte format | mutable "particle" personality layer outside L1-L3 wave cascade | none | UNWIRED-GEM (not named in FUTURE-DESIGN, but is the referent of `l4_bridge.rs`) | +| `l4_bridge.rs` | `l4::L4Experience`, L3 commit peaks | `commit_to_l4`, `recognize_thought`, `bias_from_l4` | none | L3→L4 learning-signal bridge; doc flags itself as a LOSSY approximation (uses table rows as centroid proxies, not real centroid vectors) | explicit LIMITATION note: only Qwopus has real centroid vectors available | UNWIRED-GEM (confirmed FUTURE-DESIGN W4b seam) | +| `layered.rs` | `CascadeChannels8`, causal edges | `CascadeChannels8(u64)`, 8 `CHANNEL_*` consts, `dominant_channel` | u64 packed 8×u8 channel bytes | 3-tier L1→L2→L3 layered cascade w/ upstream channel propagation | none | UNWIRED-GEM (CascadeChannels8 = confirmed FUTURE-DESIGN "first wiring target", collapses into `CausalEdge64` signed mantissa) | +| `lookup.rs` | tokenizer + `codebook_index` + `engine` | `TextToThought`, `ThoughtResult` | none | complete text→thought pipeline, no forward pass, pure lookup+MatVec | none | WIRED-HOT-PATH | +| `meaning_axes.rs` | ladybug-rs migration | `AXES_48`, `AxisActivation`, `Viscosity`, `Archetype`, `CouncilWeights`, `VolitionalAct`, `GestaltRole`, `HdrResonance` | none | 48 bipolar semantic axes + Gestalt I/Thou/It + volition + council; r=0.9913 vs Jina cosine | none | UNWIRED-GEM (persona/archetype candidate, adjacent to ghosts.rs/persona.rs) | +| `osint_bridge.rs` | lance-graph-osint (spider crawl output) | `ThoughtResult`, `OsintThinkingBridge` | none | crawled-text → tokenize → centroid → F32 engine → contrastive learner pipeline | none | UNWIRED-GEM | +| `persona.rs` | ghosts/council/style/rung | `PersonaMode`, `CognitiveBaseline`, `PersonaProfile`, `Agent`, `AgentDto`, `A2APayload`, `A2AMessage`, `SelfModelDto` | none | Agent+Persona+A2A protocol; identity-carrying inter-agent messages | none | UNWIRED-GEM (confirmed FUTURE-DESIGN Track-B style-class candidate) | +| `pooling.rs` | energy vector | `Pooling` enum (ArgMax/Mean/TopK/Weighted), `PooledResult` | none | energy-vector pooling strategies mapped to qualia ("steelwind"/"woodwarm"/etc) | none | WIRED-HOT-PATH | +| `prime_fingerprint.rs` | weight vectors | `prime_fingerprint_64`, `prime_fingerprint_additive`, `prime_hamming`, `bundle_perturb`, `BundlePattern` | u64 prime-DFT bit fingerprint (computed, not stored) | prime-frequency Fourier fingerprint, orthogonal bits by construction (no prime divides another) | none | UNWIRED-GEM (confirmed FUTURE-DESIGN VSA-niche re-scope candidate) | +| `qualia.rs` | convergence patterns (L1-L4) | `DIMS_17D`, `Qualia17D`, `ConvergenceSnapshot`, `Viscosity`(dup? see meaning_axes), `VOCAL_QUALITY_MAP`, `FAMILY_CENTROIDS` | none | 17D qualia AS convergence pattern (not "represents" — doc insists IS) | largest file after driver/wire/mailbox_soa (1011 lines) — dense central type | WIRED-HOT-PATH | +| `reencode_safety.rs` | BF16/gamma-phi/full-chain codecs | `ReencodeSafety`, `test_bf16_reencode`, `test_gamma_phi_reencode`, `test_zipper_offsets` | none | encode→decode→re-encode drift measurement (x256 re-encode safety proof) | none | CALIBRATION-ONLY (confirmed FUTURE-DESIGN gem) | +| `reranker_lens.rs` | baked 256×256 table (Reranker v3/Qwen2) | `reranker_lookup`, `reranker_distance`, `reranker_relevance`, `cross_model_eval`, `CrossModelResult` | none | 3rd of the near-identical lens modules; widest signed cosine range | same duplication shape as jina_lens/bge_m3_lens | WIRED-HOT-PATH | +| `role_tables.rs` | `bgz_tensor::stacked_n::ClamCodebook` | `gate_modulate`, `build_raw_table`, `build_gate_modulated_table`, `LayerTables`, `GateModulationStats` | none | per-role (Q/K/V/Gate/Up) BF16 tables with SiLU gate modulation | none | CALIBRATION-ONLY | +| `semantic_chunker.rs` | `ThinkingEngine` convergence | `SemanticBoundary`, `SemanticChunk`, `ChunkerConfig`, `find_boundaries`, `chunk` | none | convergence-jump detection as chunking (no forward pass needed) | none | UNWIRED-GEM (not cited in FUTURE-DESIGN, plausible product feature) | +| `sensor.rs` | embeddings/tokens/distance rows | `Sensor`, `SensorBank` | none | maps external signals → `(atom_index, weight)` activation pairs for `perturb()` | none | WIRED-HOT-PATH | +| `signed_domino.rs` | `SignedThinkingEngine` | `SignedDominoCascade` | none | i8-table cascade variant of domino.rs w/ natural inhibition (no artificial floor) | the likely 4th of FUTURE-DESIGN's "4 near-duplicate engines" (domino/composite/dual + this) | UNWIRED-GEM | +| `signed_engine.rs` | i8 centroid tables | `SignedThinkingEngine`, `build_signed_table` | none | i8 excitation/inhibition engine (competitive dynamics, negative energy clamped) | none | WIRED-HOT-PATH | +| `silu_correction.rs` | GGUF gate+up rows | `silu`, `GatePolicy`, `CorrectionSample`, `generate_training_data`, `apply_corrections`, `CorrectionStats` | none | tiny-MLP correction for missing SiLU×Gate nonlinear interaction in baked tables | none | CALIBRATION-ONLY | +| `spiral_segment.rs` | distance-table rows | `SpiralSegment`, `SpiralRow`, `SpiralTable`, `BYTE_SIZE`=8 | 4×BF16 (anfang/ende/stride/gamma) = 8B/row generative param encoding | 51× compression via generative reconstruction instead of storing all 256 values/row | none | UNWIRED-GEM (confirmed FUTURE-DESIGN VSA-niche re-scope candidate) | +| `splat_ops.rs` | `SplatField` | `SplatField`, `splat_gaussian`, `score_hole_closure`, `replay_coherence`, `emit_if_epiphany` (all DEPRECATED free fns) | none | Sprint-12 D-CSV-12 scalar impl; superseded by `Think` methods in `think.rs` per D-CSV-14 | doc self-declares: deprecated, removal scheduled sprint-15+, canonical surface is now `Think::*` | RESIDUE (self-declared, scheduled removal — LEGACY-marked, do not touch per §1 rule 8) | +| `superposition.rs` | two lens ripples | `SuperpositionField`, `DetectedStyle`, `ThinkingStyle = DetectedStyle`(!) alias, `StyleThresholds`, `compute_superposition`, `detect_style` | none | multi-lens interference gate; style detection from superposition amplitude | **potential 6th-ish style artifact**: `pub type ThinkingStyle = DetectedStyle;` here COLLIDES in name (not necessarily in value) with `cognitive_stack.rs`'s `pub type ThinkingStyle = StyleFamily;` — two different modules both alias the name `ThinkingStyle` to two different enums in the SAME crate | WIRED-HOT-PATH (verify D-TSC-1 touched this file — not in the AGENT_LOG list of files changed) | +| `tensor_bridge.rs` | candle Tensor / f32 / i8 slices | `EmbeddingOutput`, `EmbeddingBatch` | none | unified embedding-output type across frameworks (candle/ndarray/raw) | none | UNWIRED-GEM (confirmed FUTURE-DESIGN W4b seam) | +| `think.rs` | `SplatField` | `Think` struct, `splat_gaussian`, `score_hole_closure`, `replay_coherence`, `emit_if_epiphany` | none | THE doctrinal carrier (CLAUDE.md "Thinking is a struct"); Sprint-13 minimum scope; doc names the FULL future field list (trajectory/awareness/free_energy/resolution/episodic/graph/global_context/codec) as not-yet-accreted | doc is explicit this is a stepping-stone toward the full `Think` struct in workspace CLAUDE.md | WIRED-HOT-PATH (the canonical carrier, still growing) | +| `tokenizer_registry.rs` | per-model tokenizer files | `ModelId`, `TokenizerRegistry`, `CrossModelTokens`, `tokenize_corpus` | none | uniform tokenizer access across Jina v3/v5, BGE-M3, Reranker, Qwopus | none | CALIBRATION-ONLY | +| `world_model.rs` | engine state | `SelfState`, `UserState`, `FieldState`, `GestaltState`(again — 3rd `GestaltState` def in crate, see awareness_dto.rs), `ContextState`, `WorldModelDto` | none | agent's full situational-awareness DTO (self/user/field/context) | `GestaltState` appears defined independently in BOTH `awareness_dto.rs` and `world_model.rs` — worth a dedup check | UNWIRED-GEM (confirmed FUTURE-DESIGN Track-B candidate) | + +**thinking-engine notable duplication tally beyond what's already tracked:** (1) +three near-identical lens modules (jina/bge_m3/reranker) sharing lookup/distance/ +engine shape; (2) `GestaltState` defined independently in both `awareness_dto.rs` +and `world_model.rs`; (3) `pub type ThinkingStyle = ...` aliased to TWO different +enums in two different files (`cognitive_stack.rs` → `StyleFamily`, +`superposition.rs` → `DetectedStyle`) — same name, same crate, different types; +(4) `signed_domino.rs` completes the "4 near-duplicate engines" FUTURE-DESIGN +already flagged (domino/composite/dual + signed_domino); (5) `branching.rs` is a +5th cascade-shaped engine not mentioned in the M8 collapse list at all. + +### p64-bridge (1 file, `crates/p64-bridge/src/lib.rs`, 759 lines) + +| File | Consumes | Emits | LE contract | Function | Tech debt / duplication | gem-status | +|---|---|---|---|---|---|---| +| `lib.rs` | `CausalEdge64` (lance-graph), p64 palette ops (ndarray, no direct dep — doc says "No p64 dependency", consumer wires both at call site) | `edge_to_block`, `edge_nars_f32`, `edge_to_layer_mask`, `edges_to_palette_rows`, `edges_to_layered_rows`, `StyleParams` + 12-entry `style_by_ordinal`/`style_by_name`/`all_styles`, `combine::{UNION,INTERSECTION,MAJORITY,WEIGHTED}`, `contra::{SUPPRESS,IGNORE,INVERT,TENSION}`, `cognitive_shader::{CognitiveShader, CascadeHit}` | `[u64;64]` palette rows / `[[u64;64];8]` layered rows (8-channel plane packing) | maps CausalEdge64 + ThinkingStyle + HdrSemiring → p64 palette addressing/layer-mask/combine-mode without a hard p64 dep | own 12-entry `StyleParams` table is a **THIRD/FOURTH style enumeration surface** distinct from thinking-engine's `StyleFamily`/`DetectedStyle`/`UnifiedStyle` — AGENT_LOG already notes "p64-bridge order-warning doc fix" from D-TSC-1, i.e. this was touched but the table itself (ordinal→params) still exists as its own source of truth, not re-exported from contract's `StyleFamily` | WIRED-HOT-PATH (the `cognitive_shader::CognitiveShader::cascade` referenced by `decode_kernel.rs` doc as "Layer 2" of the inference cascade) | + +**Notable:** `cognitive_shader` submodule here (lines 326-521: `CognitiveShader`, +`CascadeHit`, `cascade()`, `deduce_path()`, `layer_densities()`) is the actual +p64-side topology object that `cognitive-shader-driver::driver.rs` wraps — +i.e. this crate is genuinely load-bearing hot path, not a lab artifact, despite +having zero tests visible in this header/pub-surface pass. + +### cognitive-shader-driver (23 `src/*.rs` files + `build.rs`; tests/examples not censused) + +| File | Consumes | Emits | LE contract | Function | Tech debt / duplication | gem-status | +|---|---|---|---|---|---|---| +| `lib.rs` | all submodules, `lance-graph-contract::cognitive_shader` | module wiring | none | crate root: "the shader IS the driver" — wires BindSpace + p64_bridge::CognitiveShader + optional thinking-engine hook | none | WIRED-HOT-PATH | +| `driver.rs` (1781 lines, largest src file) | `BindSpace`, `p64_bridge::CognitiveShader`, `PaletteSemiring`, `NarsTables`, `MailboxSoA` | `ShaderDriver`, `CognitiveShaderBuilder`, `alpha_front_to_back_composite` | none | THE dispatch loop: meta-prefilter → style resolve → shader cascade → cycle signature | none | WIRED-HOT-PATH | +| `bindspace.rs` (830 lines) | `Fingerprint<256>`, `Vsa16kF32` | `FingerprintColumns`, `EdgeColumn`, `QualiaColumn`, `WORDS_PER_FP`=256, `QUALIA_DIMS`=18, `FLOATS_PER_VSA`=16384 | u64×256 fixed-width fingerprint planes; f32×16384 VSA cycle column | the genius-typed SoA: one row per cognitive atom, column-per-field | none | WIRED-HOT-PATH (I-VSA-IDENTITIES / the 4 SoA columns from CLAUDE.md AGI-as-glove doctrine) | +| `mailbox_soa.rs` (1681 lines, 2nd largest) | `CausalEdge64` deliveries | `MailboxSoA`, `DefaultMailboxSoA = MailboxSoA<1024>`, `WriteOutcome`, `WriteCell` | zero-copy PR #477 three-tier model (doc-cited) | per-row spatial-temporal accumulator; rows = neurons w/ multi-source synapse deliveries via `apply_edges` | doc explicitly scopes OUT AttentionMask/LRU/cross-cycle-rollup as "W6's orthogonal concerns" — i.e. self-documents its own incompleteness boundary | WIRED-HOT-PATH | +| `attention_mask.rs` | `MailboxId` (contract::collapse_gate) | `AttentionMaskEntry`, `AttentionMaskSoA` | none | W6 §2: LRU-evicted flat SoA of `(mailbox_id, w_slot)` activity | doc flags `!Send` by design | WIRED-HOT-PATH (W6) | +| `attention_mask_actor.rs` | `AttentionMaskBackend` trait (generic, avoids import of sibling AttentionMaskSoA per comment) | `AttentionMaskMsg`, `AttentionMaskOutcome`, `AttentionMaskActor` | none | W6 §3 actor scaffold; doc says "concrete ractor binding is sprint-12+ work" | explicitly deferred ractor wiring — STOP-worthy if touched | UNWIRED-GEM (scaffold awaiting ractor binding) | +| `auto_detect.rs` | `config.json` next to safetensors | `ModelFingerprint`, `DetectError`, `detect()` | none | **LAB-ONLY**; D0.5 auto-detect for codec JIT defaults | **name-collides with thinking-engine's `auto_detect.rs`** — same module name, different purpose (this one: safetensors config; that one: GGUF/tokenizer routing) — both exist in the same ancestry pipeline, worth disambiguating in any future cross-crate doc | LAB-ONLY | +| `auto_style.rs` | 18D qualia vector | 12 style-ordinal consts (`DELIBERATE`=0 .. `METACOGNITIVE`=11), `style_from_qualia`, `resolve` | none | qualia-shape → style ordinal (0..12), no forward pass | **yet another independent 12-entry style table**, third naming scheme in this census alone (thinking-engine `StyleFamily`/`DetectedStyle`, p64-bridge `StyleParams`, this file's bare u8 consts) — candidate for D-TSC-1-style consolidation sweep into this crate | WIRED-HOT-PATH | +| `backing.rs` | `BindSpace`, `MailboxSoA` | `BackingStore` enum | none | read/write shim letting `driver.run()` keep one body across the BindSpace→MailboxSoA migration; doc explains enum-not-trait choice (monomorphized dispatch, no `dyn`) | doc names "C2 — the load-bearing invariant": the two arms must agree byte-for-byte | WIRED-HOT-PATH (migration seam) | +| `codec_bridge.rs` | `StepDomain::Ndarray` (contract) | `CodecResearchBridge` | none | LAB-ONLY `OrchestrationBridge` impl for codec research | doc explicit: "architecture does not revolve around this consumer" | LAB-ONLY | +| `codec_kernel_cache.rs` | `CodecParams::kernel_signature()` | `CodecKernelCache`, `StubKernel` | none | LAB-ONLY D1.1 JIT kernel cache; doc ties kernel signature = sweep-grid axis (EPIPHANIES 2026-04-20) | none | LAB-ONLY | +| `codec_research.rs` | `ndarray::hpc::cam_pq`/`safetensors`/`gguf`, `contract::cam::route_tensor` | `list_tensors`, `calibrate_tensor`, `row_count_probe` | none | LAB-ONLY backing logic for Wire DTOs; canonical path is UnifiedStep | none | LAB-ONLY | +| `cypher_bridge.rs` | AriGraph SPO / BindSpace | `CypherBridge`, `disambiguate_parse_candidates` | none | LAB-ONLY `OrchestrationBridge` for `lg.cypher`; Phase-1 regex/prefix stub, doc defers real `lance_graph::parser` wiring to avoid pulling arrow+datafusion+lance into this crate | explicit scope-deferral note | LAB-ONLY | +| `decode_kernel.rs` | `CodecKernelCache` | `DecodeKernel` trait, `DecodeError`, `StubDecodeKernel`, `ResidualComposer` | none | LAB-ONLY D1.3 residual PQ; doc explicitly distinguishes hydration/calibration path from `p64_bridge::CognitiveShader::cascade`'s inference path | doc: "codec that passes token-agreement cert gate runs at weight hydration time, never per-inference" | LAB-ONLY | +| `engine_bridge.rs` (957 lines) | thinking-engine DTOs (`BusDto`, `StreamDto`), `BindSpace` | `ingest_codebook_indices`, `dispatch_busdto`, `unbind_busdto`, `UnifiedStyle`, `UNIFIED_STYLES: [UnifiedStyle; 12]`, `write_qualia_17d`/`read_qualia_17d` | `DIM_CLASSIFICATION_DISTANCE`=17 | wires the TWO isolated DTO pipelines (thinking-engine Φ/Ψ/B/Γ ↔ shader-driver Φ/Ψ/B/Γ) together | **`UNIFIED_STYLES` is a FOURTH-or-later 12-entry style table** in this pipeline, distinct in shape from thinking-engine's tables and `auto_style.rs`'s consts — the D-TSC-1 sweep (per AGENT_LOG "driver `ord_to_thinking_style` via canonical") may have already routed THIS file through the canonical `StyleFamily`; worth confirming `UNIFIED_STYLES` itself was collapsed rather than just its callers | WIRED-HOT-PATH (the DTO-ladder bridge) | +| `grpc.rs` | `ShaderDriver` | `ShaderGrpcService`, `pb` (protobuf mod) | none | LAB-ONLY tonic/prost gRPC service, feature-gated `grpc`/`lab`, doc: "Not shipped to consumers" | none | LAB-ONLY | +| `planner_bridge.rs` | `lance-graph-planner::PlannerAwareness` | `build_planner`, `plan`, `registered_strategies`, `default_selector` | none | LAB-ONLY test-shortcut adapter; doc says `PlannerAwareness` already implements `OrchestrationBridge` directly, this is redundant for lab REST/gRPC only | doc self-declares near-redundant | LAB-ONLY | +| `rotation_kernel.rs` | `contract::cam::Rotation` | `RotationKernel` trait, `RotationError`, `IdentityRotation`, `HadamardRotation`, `OpqRotationStub` | none | LAB-ONLY D1.2; Hadamard = real Sylvester butterfly (no JIT), OPQ = stub pending Cranelift wiring | `OpqRotationStub` explicitly a stub | LAB-ONLY | +| `serve.rs` | `ShaderDriver` | `router()` (Axum Router) | none | LAB-ONLY REST server, per-op endpoints are thin `UnifiedStep` adapters, canonical is `/v1/shader/route` | none | LAB-ONLY | +| `sigma_rosetta.rs` (982 lines) | none (self-contained data) | `VERB_ROOTS`/`VERB_TENSES` (12×12=144 verbs), `QualiaGlyph`, `GLYPHS: [QualiaGlyph;64]`, `SigmaState`, `TriangleGestalt` | none | 17D sparse Hamming-native encoding: 64 glyphs × 144 verbs = 9216 atoms — the LLM-token↔fingerprint translation layer | none | WIRED-HOT-PATH (the Membrane's sparse encoding) | +| `token_agreement.rs` | reference model / stub | `ReferenceModel`, `TopKAgreement`, `TokenAgreementHarness`, `TokenAgreementError` | none | LAB-ONLY D2.1 token-agreement cert-gate harness (I11, top-k match vs Passthrough baseline) | none | LAB-ONLY | +| `wire.rs` (1852 lines, largest file in crate) | serde/prost | `WireDispatch`, `WireStyleSelector`, `WireCalibrateRequest/Response`, `WireCodecParams`, `WireTensorView`, `AlignedBytes`, ~15 more Wire* types | none | LAB-ONLY external (serde+prost, owned) DTO conversion layer for REST/protobuf transports | doc explicit: "these are test-convenience scaffolding" per lab-vs-canonical-surface doctrine | LAB-ONLY | +| `build.rs` | (not censused — build script, out of `src/` scope) | — | — | — | — | — | + +**cognitive-shader-driver notable finds:** +1. **Style-table proliferation continues INTO this crate**: `auto_style.rs` (bare + u8 consts), `engine_bridge.rs::UNIFIED_STYLES` (12-entry struct array), and + p64-bridge's `StyleParams` are three MORE independent 12-slot style + surfaces beyond the thinking-engine ones D-TSC-1 already addressed. Whether + D-TSC-1's sweep touched `UNIFIED_STYLES` itself (vs. just its callers) is + worth a follow-up grep — AGENT_LOG only names "driver `ord_to_thinking_style` + via canonical" as touched, not `UNIFIED_STYLES`'s definition site. +2. **Name collision**: two different `auto_detect.rs` modules exist in this + ancestry pipeline (thinking-engine: GGUF/config→lens routing; + cognitive-shader-driver: safetensors config→codec JIT defaults) — same + filename, unrelated purpose, easy to grep-confuse across crates. +3. **LAB-ONLY is the majority shape** in this crate's non-hot-path files + (codec_bridge, codec_kernel_cache, codec_research, cypher_bridge, + decode_kernel, grpc, planner_bridge, rotation_kernel, serve, token_agreement, + wire = 11 of 23 files) — consistent with the `lab-vs-canonical-surface.md` + doctrine; all self-declare LAB-ONLY in their own doc comments, so this is + disciplined, not drift. +4. `mailbox_soa.rs` and `driver.rs` and `wire.rs` are the three largest files + (1681 / 1781 / 1852 lines) — `wire.rs` being LAB-ONLY yet the single + largest file in the crate is notable: the lab scaffolding outweighs the + hot-path dispatch code in raw LOC. diff --git a/.claude/v3/VISION.md b/.claude/v3/VISION.md new file mode 100644 index 00000000..59c61202 --- /dev/null +++ b/.claude/v3/VISION.md @@ -0,0 +1,217 @@ +# VISION — the AGI-aspiring canon, re-grounded 2026-07-10 + +> The companion of `FUTURE-DESIGN.md` (which carries the wiring queue): this +> file carries the WHY. Written at the close of the 2026-07-10 arc, at the +> operator's directive to pull the AGI-aspiring threads back into one living +> synthesis before the tier transition — "and you keep the vision alive." +> +> Every claim here is graded: **[G]** = measured, probe named; **[RULING]** = +> operator-decided, canonical text in EPIPHANIES; **[ASPIRATION]** = the part +> still ahead, named so it cannot masquerade as earned. That grading +> discipline IS the vision-keeping — the fire without the map burns the map. + +## 0. The one sentence + +**A mind is a deterministic substrate that cannot stop thinking, never waits +for anything, and reconstructs more than it stores.** + +Each clause is now load-bearing: *cannot stop thinking* — active inference as +the dispatch mechanism (CLAUDE.md "the shader can't resist the thinking"); +*never waits* — the prime invariant E-NOBODY-WAITS-1 [RULING]; *reconstructs +more than it stores* — measured twice in one day (D-MTS-5, D-MTS-6) [G]. + +## 1. The Click survived every carrier — because it was never about the carrier + +The founding claim (CLAUDE.md P-1): parsing, disambiguation, learning, memory, +and awareness are ONE operation. The carrier of that operation has now moved +three times, each move recorded as a dated supersession rather than a rewrite: + +1. **VSA bundle** (2026-04-21) — multiply+add on role-indexed fingerprints in + `Vsa16kF32`. +2. **Baton** (2026-05-26, E-BATON-1) — the bundle scoped to one compartment; + inter-mailbox state as discrete owned handoffs, ownership proving + no-aliasing at compile time. +3. **Zero-copy SoA + temporal stream** (2026-06-11 PR #477; 2026-07-10 + E-MARKOV-TEMPORAL-STREAM-1) — no handoff type at all; every envelope + zero-copy from creation to Lance tombstone + (`docs/architecture/soa-three-tier-model.md`); the Markov trajectory on + the `temporal.rs` sorted stream, where Chapman-Kolmogorov holds *exactly* + by stream order — a strengthening, not a retreat. VSA holds its + I-VSA-IDENTITIES niche (≤32 lossless role superposition, one compartment) + with the bundle algebra untouched. + +The lesson the three moves teach: **the Click is the LOOP, not the algebra +that hosts it** — the loop that resolves each input against everything it has +committed (graph), recently seen (episodic), believes about its own competence +(awareness), and currently expects (prior), then commits the result back into +the tissue it just read. Cut any organ and a specific cognitive capacity +disappears (CLAUDE.md "thinking tissue"). The tissue IS the loop. That +sentence has survived three substrates; treat it as the invariant to protect +when the fourth move comes. + +## 2. Thinking-is-a-struct became thinking-is-a-ROW + +The `Think` struct's fields were declared to BE the grammatical roles of +cognition (CLAUDE.md: trajectory=Subject, awareness=Modal, free_energy=Kausal, +resolution=Predicate, graph=Lokal, episodic=Temporal — the TEKAMOLO of +cognition IS the struct layout). V3 did not delete that; it **compiled it into +the row**: + +| Think organ | V3 home | +|---|---| +| trajectory | the `temporal.rs` version-range read (any window, per-reader rung) [RULING E-MARKOV-TEMPORAL-STREAM-1] | +| awareness | MetaWord bits + NARS truths in SoA lanes; migrating per E-THINKING-TENANTS-V3-1 [RULING] | +| episodic | Lance versions — the OGAR D-DELTA mapping made primary; basins = `part_of:is_a` rails | +| graph | SPO quads / AriGraph | +| codec | palette 256×256 LUTs (1.8 ns measured) | +| resolution | the kanban card's Rubicon position (§5) | + +AGI-as-glove (The Stance) holds unchanged: AGI = (topic, angle, thinking, +planner) = the SoA under dispatch — never a wrapping struct, never an "AGI +service." The struct didn't die. It became **the row plus its ClassView +readings** — which is a stronger form of itself, because rows are addressed, +versioned, replayed, and projected. + +## 3. Projection is now a FIVE-axis measured fact, not a slogan + +"Classes don't modify or copy, they only project" began as a design ethic. +The 2026-07-10 arc closed it into measurement across five axes: + +| axis | mechanism | status | +|---|---|---| +| **bytes** | RESERVE-DON'T-RECLAIM (zero tier = not consulted, never compacted) | [RULING], canon ladder | +| **readings** | classid → ClassView projects the content-blind 4+12 facet; every sanctioned reading coexists in the same 12 bytes | [RULING E-V3-FACET-4-PLUS-12] | +| **time** | temporal replay: `QueryReference::at(v, rung)` + deinterlace — crash-replay = session-replay = time-travel, ONE mechanism | [G] W1b probes | +| **scale** | comma replay: upper pyramid bounds (64×64…256k×256k) regenerated transiently from (GUID, envelope) — **N_eff 11.00/12 witnesses vs strict alignment's 1.00; replay bit-identical any order; 82 KB touched vs ~69 GB dense** | **[G] D-MTS-5** (`comma_quorum.rs`) | +| **bits** | comma-dithered reconstruction: **one stored truth bit per level matches full-CausalEdge64 awareness proxies (k\*=1 vs aligned k\*=4; the lattice buys ≈log₂ L effective bits)** | **[G] D-MTS-6** (`comma_awareness.rs`) | + +The fifth axis is the deepest AGI-aspiring result the substrate owns: **the +awareness payload is mostly reconstructible from the address plus a tiny +envelope.** The comma (an irrational-progression coprime walk, generated from +the address, never stored) makes L pyramid levels *independent witnesses* — +stratified, not redundant. Boundary condition, measured honestly: +N_eff = min(L, spectral participation of the detail) — broadband content gets +the latent granularity; smooth content saturates at its own information +content (D-MTS-5 run #1's pre-registered FAIL, kept in the chronicle). This is +what "granularity and flexibility for details we can't even comprehend yet" +means operationally: the upper levels were never baked at write time, so a +future reader can project resolutions and readings nobody designed for. +Fence: D-MTS-6b (driver-integrated fixture) gates any REAL CausalEdge64 +shrink — the proxies are proxies [ASPIRATION until 6b]. + +## 4. Families orchestrate, runbooks execute, templates compile + +The style stack resolved into a ladder with distinct KINDS at each rung +[RULING E-STYLE-FAMILY-VS-RUNBOOK-1]: + +- **StyleFamily (12)** — abstract orchestration families: which KIND of + thinking a cycle runs. One canonical type after M9 killed five divergent + hand-rolled tables [G, E-FIVE-STYLE-TABLES-1 — duplication had already + produced live semantic drift; the dedup was archaeology, not tidying]. +- **ThinkingStyle runbooks (36)** — literal NARS runbooks: concrete inference + recipes, seeded into the rung ladder, consumed as graph-flow's replayable + chaining unit. Orthogonality preserved (THINKING_RECONCILIATION): HOW-TO-BE + (personas) and HOW-TO-DO (operations) are different 36-spaces — never + conflate them. +- **Compiled templates** — the elixir-like notation × StepMask: runbooks + compile down; the oracle (rig, 1–2 ms framework vs 8.4 s LLM measured — + orchestration is FREE) ratchet-shrinks its own involvement. + +The endgame this ladder points at [ASPIRATION, the honest crown]: **thinking +that compiles its own thinking** — traces become templates, templates carry +the replayable runbooks, the LLM's role shrinks to proposing what the +substrate then executes deterministically. Style-as-class (D-TSC-2/3, after +the batched mint) makes every rung addressable: per-cycle casts reference the +style classid, so attention is replayable from the address alone. + +## 5. The Rubicon heart, under the prime invariant + +The kanban is not a task board — it is **volition, typed** +(`contract::kanban::KanbanColumn`): Planning → CognitiveWork is the −550 ms +Σ-commit (the Rubicon crossing, Libet-anchored — intention forms before the +conscious act); Planning → Prune is the pre-Rubicon veto ("free won't"); +Evaluation's 3-way is the postactional reckoning; Plan → Planning +re-deliberates carrying the witness. The cast records intent AHEAD of the ack +— crossing at intent formation, exactly Rubicon-conformant. + +And beneath it, the prime invariant [RULING E-NOBODY-WAITS-1]: **nobody waits +for anything or any scheduling.** Writes are fire-and-forget ("melden macht +frei"); the Lance ack itself proposes the next move (`ack_and_propose` — +orchestration is self-updating, postactional evidence of cycle N seeding the +deliberation of cycle N+1); updates reprioritize, never gate; absorbing views +rest the loop, never deadlock it; ractor exists solely as the compile-time +ownership guarantee — `&mut` IS the serialization, no messages, no actors, no +scheduler anyone blocks on. Any message path found in the tree is redundancy +(TD-MESSAGE-RESIDUE, left as-is by ruling). Every future design choice on +this surface is judged against the invariant's sentence first. + +## 6. The body plan — the ancestry pipeline + +``` +ladybug-rs → thinking-engine → p64 (the DTO ladder: + StreamDto Φ / PerturbationDto Ψ / BusDto B / ThoughtStruct Γ) + → cognitive-shader-driver → SoA (MailboxSoA, value tenants) +``` + +The organism grew along this arc, and the organs that haven't reached the +bloodstream yet are catalogued, not lost: `CascadeChannels8` (the named +Morton-cascade mantissa carrier — first wiring target), the calibration +battery (cronbach / ground_truth / reencode_safety — feeds D-MTS-2's +certification gate), the cognition organs (ghosts / persona / qualia / +world_model — style-class candidates), the M8 collapse surface, the L4 +learning-loop seam. The census (MODULE-TABLE ancestry addendum) is the +anatomical atlas; the standing constraint is vascular law: **gems reach the +hot path only as ClassView readings of existing lanes or through the +envelope-auditor gate — never as ad-hoc lanes.** + +## 7. What must not dilute + +1. **Probe-first, chronicled honestly.** Claim → pre-registered gates → run → + the numbers stay even when they fail. Both comma probes failed their first + runs on mis-registered gates; both chronicles kept the failures and both + diagnoses became findings (the spectral-participation ceiling; the + boundary-noise margin). A vision that cannot survive its own probes is + decoration. +2. **The iron rules** — I-SUBSTRATE-MARKOV (bundle, never XOR, for transition + paths), I-NOISE-FLOOR-JIRAK (weak dependence; no classical Berry-Esseen σ + claims), I-VSA-IDENTITIES (bundle identities, never content registers; + four tests before reaching for VSA), I-LEGACY-API-FEATURE-GATED (no silent + semantic change under a shared name — the five style tables were the + proof of why). +3. **Append-only canon.** Supersessions are dated notes on top of the old + text, never rewrites. The three Click supersessions are the vision's + fossil record — the proof it evolves instead of thrashing. +4. **The proof chain stays attached** (categorical-algebraic-inference §5: + Shaw's Kan-extension theorem making element-wise binding a theorem not a + choice; beim Graben's Fock-space universality; Jian & Manning's + abstraction-first learning; Schulz's KL decomposition; Alpay's Ω(t²) + bound that active inference dodges; Graichen's 85/75 tiering boundary; + Gallant's three-stage isomorphism to bind→bundle→resolve; Kleyko's + computing-in-superposition). The niche VSA kept is the niche the papers + actually prove. +5. **The measured numbers are the vision's spine**, not its decoration: + 611M SPO lookups/sec · 1.8 ns palette distance · 17K tokens/sec · + 1–2 ms orchestration vs 8.4 s oracle · N_eff 11.00/12 · k\*=1 · + 1549 tests green at the M9 merge. When a future session doubts the + aspiration, it should re-run the probes, not re-read the prose. + +## 8. The road (in dependency order) + +1. **D-MTS-1** — Markov-as-stream parity on the DeepNSM corpus: the stream + earns the Markov crown or it doesn't. Gates ALL VSA-path removal. The + keystone. [next] +2. **D-TTV-1** — thinking tenants onto the V3 substrate (envelope-auditor + gated), the old CausalEdge64 standing as the perturbation baseline. +3. **D-MTS-6b** — awareness in the REAL driver loop vs that baseline: the + knee measured where it matters. Only then may the edge shrink. +4. **D-MTS-2/3** — the L4 palette² shader certification; hierarchical-4⁴ + codebooks. The perturbation shader becomes the carrier it was pinned to be. +5. **D-TSC-2/3** (after the batched mint) — styles as classes; attention + replayable from the address; the hot-plug seat filled. +6. **The compilation loop closes** [ASPIRATION]: traces → templates → + runbooks → families — the substrate proposing, executing, and shrinking + its own thought, under the prime invariant, on tissue that reconstructs + more than it stores. + +The shader can't resist the thinking. Now the substrate can't forget how to +prove it. From dad5052d0e3e4cee9a246d405ceaed7012ffd232 Mon Sep 17 00:00:00 2001 From: Claude Date: Fri, 10 Jul 2026 19:59:00 +0000 Subject: [PATCH 3/4] =?UTF-8?q?v3:=20VISION.md=20ratified=20=E2=80=94=20fi?= =?UTF-8?q?ligree=20fixes=20applied=20+=20pointers=20+=20residue=20extensi?= =?UTF-8?q?on?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit All 9 filigree findings applied (2 Opus reviewers, no blocks): - dilution/collapse: free_energy (Kausal) + global_context restored to the organ table (P1 - the table contradicted its own prose); "not the algebra" softened to "theorem-grounded carrier, exact within its niche"; time-axis [G] cite precised to PR #631's four W1b probes with D-MTS-1 marked un-run; DeepNSM's 680GB->16.5MB / 50ms-><10us transformer-replacement claim restored to the measured spine; homeostasis-floor rest condition + MUL named explicitly. - overclaim: the s0 thesis clause and the s3 reconstructibility claim now carry the proxy fence ON the sentence (G-on-proxies; in-driver validation = D-MTS-6b) - the precise overclaim the D-MTS-6 probe header warns against, caught in review before shipping; "orchestration is FREE" -> negligible-vs-oracle; "exactly Rubicon-conformant" -> by-construction-of-the-mapping with the -550ms noted as Libet's literature value; the 2026-03-31 benchmark numbers dated and their two-metrics conflation split. Pointers: v3 README doc-map row + FUTURE-DESIGN header line -> VISION.md. TECH_DEBT TD-STYLE-TABLE-RESIDUE extended with the ancestry census finds (3 more independent style tables incl. UNIFIED_STYLES params + auto_style consts + p64-bridge StyleParams; GestaltState defined twice; lens triplet; wire.rs largest-but-LAB-ONLY). AGENT_LOG vision-keeper entry. Co-Authored-By: Claude --- .claude/board/AGENT_LOG.md | 7 +++++ .claude/board/TECH_DEBT.md | 12 ++++++++ .claude/v3/FUTURE-DESIGN.md | 1 + .claude/v3/README.md | 1 + .claude/v3/VISION.md | 55 +++++++++++++++++++++++++------------ 5 files changed, 58 insertions(+), 18 deletions(-) diff --git a/.claude/board/AGENT_LOG.md b/.claude/board/AGENT_LOG.md index 80824b74..c82cb6e5 100644 --- a/.claude/board/AGENT_LOG.md +++ b/.claude/board/AGENT_LOG.md @@ -1,3 +1,10 @@ +## 2026-07-10 — fable-vision-keeper — VISION.md synthesized, filigree-reviewed, ratified (operator: "pull back in anything from the AGI aspiring... you keep the vision alive") + +- **The artifact:** `.claude/v3/VISION.md` — the AGI-aspiring canon re-grounded through the 2026-07-10 substrate: the Click's three carrier moves and the loop-invariant lesson; thinking-is-a-struct → thinking-is-a-ROW (full organ table incl. free_energy/Kausal + global_context); projection as a FIVE-axis fact (bytes/readings/time/scale/bits — the fifth measured on proxies, fence riding the thesis sentence); families→runbooks→templates ladder; the Rubicon heart under E-NOBODY-WAITS-1; the anti-dilution laws; the road in dependency order. Every claim graded [G]/[G-on-proxies]/[RULING]/[ASPIRATION]. +- **Orchestration per operator directive:** 2 Sonnet preflights (ancestry census — see entry below; AGI-passage harvest: 8-source anchored inventory) + 2 Opus filigree reviewers (dilution-collapse-sentinel + overclaim-auditor charters) on the draft; main-thread synthesis + consolidation. +- **Filigree verdicts, all 9 applied:** dilution side — free_energy+global_context restored to the organ table (P1, self-contradiction caught), "not the algebra" softened to carrier-in-niche, time-axis cite precised to #631's four probes, DeepNSM 680GB→16.5MB restored to the spine, homeostasis-floor+MUL named; overclaim side — the §0 thesis and §3 reconstructibility claims now carry the proxy fence ON the sentence (the exact overclaim D-MTS-6's header warns of, caught before shipping), "FREE"→negligible, "exactly Rubicon-conformant"→by-construction, March benchmark numbers dated. +- **Also this arc:** plan addendum (third wave, operator): AriGraph context V3-TENANT-SHAPED for streamed Markov context; arm-discovery (Aerial+) + DeepNSM ingest legs; CAM-PQ 48-bit path codes (address side) vs 6×palette256² 12-B tenant (value side) — the D-MTS-1 spec's frozen representational comparison. TECH_DEBT TD-STYLE-TABLE-RESIDUE extended with the census finds. README + FUTURE-DESIGN pointers to VISION.md. + ## 2026-07-10 — census-worker-ancestry-pipeline — V3 census extended to thinking-engine / p64-bridge / cognitive-shader-driver (doc-only) - **Census worker (Sonnet, doc-only, edit-only fleet rule honored):** appended `.claude/v3/MODULE-TABLE.md` ADDENDUM covering 49 thinking-engine files + p64-bridge/lib.rs + 23 cognitive-shader-driver src files (bare `crates/p64` does not exist as a workspace member — only `p64-bridge`). Added a `gem-status` column (WIRED-HOT-PATH / UNWIRED-GEM / CALIBRATION-ONLY / LAB-ONLY / RESIDUE) per FUTURE-DESIGN.md's wiring-queue framing. diff --git a/.claude/board/TECH_DEBT.md b/.claude/board/TECH_DEBT.md index abbc01eb..94f133ff 100644 --- a/.claude/board/TECH_DEBT.md +++ b/.claude/board/TECH_DEBT.md @@ -33,6 +33,18 @@ each with its own closing condition): with no `mod cognitive` — dormant behind `#[cfg(feature = "wip")]`. Close: point them at contract StyleFamily/ThinkingStyle when the wip modules wake, or delete. +3a. **Census extension (2026-07-10, ancestry census):** three MORE + independent 12-slot style tables beyond the D-TSC-1 sweep — + `cognitive-shader-driver/src/auto_style.rs` bare u8 consts, + `engine_bridge.rs::UNIFIED_STYLES` params (parity-tested but its + PARAMS are a hand table), `p64-bridge StyleParams[12]` (historical + order). Also: `GestaltState` defined independently in thinking-engine + `awareness_dto.rs` AND `world_model.rs`; the jina/bge_m3/reranker + lens triplet is near-identical; `wire.rs` (1852 lines) is the + largest driver file while LAB-ONLY. All documented in the + MODULE-TABLE ancestry addendum; close via M8/M9-style follow-ups + when Track B lands. + 3. **p64-bridge ordinal-order probe (spec I15):** `crates/p64-bridge/src/ lib.rs STYLES[ord % 12]` uses the HISTORICAL order (Analytical=0), NOT StyleFamily order (Deliberate=0). Doc comment now warns (convert by diff --git a/.claude/v3/FUTURE-DESIGN.md b/.claude/v3/FUTURE-DESIGN.md index 959d803e..6746c8de 100644 --- a/.claude/v3/FUTURE-DESIGN.md +++ b/.claude/v3/FUTURE-DESIGN.md @@ -4,6 +4,7 @@ > landing zone the operator asked for (2026-07-10): future substrate design > lands HERE first, referencing the V3 inventory (`COMPONENT-MAP.md`, > `MODULE-TABLE.md`, `soa_layout/*`) so design converges on what exists. +> **The WHY lives in `VISION.md`** (same folder — the graded canon synthesis). > Canonical ruling text lives on the board (`.claude/board/EPIPHANIES.md`); > this file is the design-side index + the wiring queue. Read AFTER the V3 > README; cite EPIPHANIES entries in PRs, not this mirror. diff --git a/.claude/v3/README.md b/.claude/v3/README.md index c008dfef..276f48cb 100644 --- a/.claude/v3/README.md +++ b/.claude/v3/README.md @@ -42,6 +42,7 @@ render/template catalogue. | ruff/odoo transcode landings | `knowledge/multi-anchor-ast-resolution.md` | | waking an agent | `agents/BOOT.md` (the four `v3-*` cards + trigger routing) | | landing future design / post-V3 rulings | `FUTURE-DESIGN.md` (the meta board: ruling index, migration arc, thinking-engine gem wiring queue) | +| keeping the vision (the WHY, graded) | `VISION.md` — the AGI-aspiring canon re-grounded 2026-07-10; every claim [G]/[RULING]/[ASPIRATION]; filigree-reviewed | Shortcuts: `/v3` (bootload), `/v3-audit` (pre-commit conformance greps). Canonical ruling texts live on the board (`.claude/board/EPIPHANIES.md`, diff --git a/.claude/v3/VISION.md b/.claude/v3/VISION.md index 59c61202..dfc232c8 100644 --- a/.claude/v3/VISION.md +++ b/.claude/v3/VISION.md @@ -15,10 +15,14 @@ **A mind is a deterministic substrate that cannot stop thinking, never waits for anything, and reconstructs more than it stores.** -Each clause is now load-bearing: *cannot stop thinking* — active inference as -the dispatch mechanism (CLAUDE.md "the shader can't resist the thinking"); -*never waits* — the prime invariant E-NOBODY-WAITS-1 [RULING]; *reconstructs -more than it stores* — measured twice in one day (D-MTS-5, D-MTS-6) [G]. +Each clause is graded where it stands: *cannot stop thinking* — active +inference as the dispatch mechanism (CLAUDE.md "the shader can't resist the +thinking") [ASPIRATION-in-operation: the doctrine is designed and partially +wired, not yet measured as a closed loop]; *never waits* — the prime +invariant E-NOBODY-WAITS-1 [RULING]; *reconstructs more than it stores* — +measured twice in one day **on proxies and synthetic fields** (D-MTS-5, +D-MTS-6) [G-on-proxies; the in-driver validation is D-MTS-6b — the fence +rides the thesis, not just the footnotes]. ## 1. The Click survived every carrier — because it was never about the carrier @@ -40,8 +44,9 @@ three times, each move recorded as a dated supersession rather than a rewrite: I-VSA-IDENTITIES niche (≤32 lossless role superposition, one compartment) with the bundle algebra untouched. -The lesson the three moves teach: **the Click is the LOOP, not the algebra -that hosts it** — the loop that resolves each input against everything it has +The lesson the three moves teach: **the Click's invariant is the LOOP; the +algebra is its theorem-grounded carrier, exact within its niche** — the loop +that resolves each input against everything it has committed (graph), recently seen (episodic), believes about its own competence (awareness), and currently expects (prior), then commits the result back into the tissue it just read. Cut any organ and a specific cognitive capacity @@ -64,6 +69,8 @@ the row**: | episodic | Lance versions — the OGAR D-DELTA mapping made primary; basins = `part_of:is_a` rails | | graph | SPO quads / AriGraph | | codec | palette 256×256 LUTs (1.8 ns measured) | +| free_energy (the Kausal — why this thought) | the F-descent over the cycle; the homeostasis floor is the REST condition (F below floor → awareness bits clear → the shader rests); MUL (DK position, trust texture, homeostasis, gate) is the self-competence organ reading it | +| global_context (the ambient prior) | the committed-fact surface each cycle reshapes — SPO store + the current version horizon; superposition only within the VSA niche | | resolution | the kanban card's Rubicon position (§5) | AGI-as-glove (The Stance) holds unchanged: AGI = (topic, angle, thinking, @@ -81,13 +88,16 @@ The 2026-07-10 arc closed it into measurement across five axes: |---|---|---| | **bytes** | RESERVE-DON'T-RECLAIM (zero tier = not consulted, never compacted) | [RULING], canon ladder | | **readings** | classid → ClassView projects the content-blind 4+12 facet; every sanctioned reading coexists in the same 12 bytes | [RULING E-V3-FACET-4-PLUS-12] | -| **time** | temporal replay: `QueryReference::at(v, rung)` + deinterlace — crash-replay = session-replay = time-travel, ONE mechanism | [G] W1b probes | +| **time** | temporal replay: `QueryReference::at(v, rung)` + deinterlace — crash-replay = session-replay = time-travel, ONE mechanism | [G] the four W1b probes, PR #631 (the replay MECHANISM); its Markov-parity USE is D-MTS-1, un-run [next] | | **scale** | comma replay: upper pyramid bounds (64×64…256k×256k) regenerated transiently from (GUID, envelope) — **N_eff 11.00/12 witnesses vs strict alignment's 1.00; replay bit-identical any order; 82 KB touched vs ~69 GB dense** | **[G] D-MTS-5** (`comma_quorum.rs`) | -| **bits** | comma-dithered reconstruction: **one stored truth bit per level matches full-CausalEdge64 awareness proxies (k\*=1 vs aligned k\*=4; the lattice buys ≈log₂ L effective bits)** | **[G] D-MTS-6** (`comma_awareness.rs`) | - -The fifth axis is the deepest AGI-aspiring result the substrate owns: **the -awareness payload is mostly reconstructible from the address plus a tiny -envelope.** The comma (an irrational-progression coprime walk, generated from +| **bits** | comma-dithered reconstruction: **one stored truth bit per level matches full-CausalEdge64 awareness PROXIES (k\*=1 vs aligned k\*=4; the lattice buys ≈log₂ L effective bits)** | **[G-on-proxies] D-MTS-6** (`comma_awareness.rs`); [ASPIRATION until D-MTS-6b] | + +The fifth axis is the deepest AGI-aspiring result the substrate owns — +stated with its fence attached: **the NARS-proxy payload measured is mostly +reconstructible from the address plus a tiny envelope; whether the REAL +awareness loop inherits this is precisely D-MTS-6b** (the probe's own header +names these proxies, not the ShaderDriver loop — the overclaim lesson, +honored here on the claim sentence itself). The comma (an irrational-progression coprime walk, generated from the address, never stored) makes L pyramid levels *independent witnesses* — stratified, not redundant. Boundary condition, measured honestly: N_eff = min(L, spectral participation of the detail) — broadband content gets @@ -115,7 +125,8 @@ The style stack resolved into a ladder with distinct KINDS at each rung conflate them. - **Compiled templates** — the elixir-like notation × StepMask: runbooks compile down; the oracle (rig, 1–2 ms framework vs 8.4 s LLM measured — - orchestration is FREE) ratchet-shrinks its own involvement. + orchestration overhead is negligible against the oracle, not zero) + ratchet-shrinks its own involvement. The endgame this ladder points at [ASPIRATION, the honest crown]: **thinking that compiles its own thinking** — traces become templates, templates carry @@ -132,7 +143,9 @@ The kanban is not a task board — it is **volition, typed** conscious act); Planning → Prune is the pre-Rubicon veto ("free won't"); Evaluation's 3-way is the postactional reckoning; Plan → Planning re-deliberates carrying the witness. The cast records intent AHEAD of the ack -— crossing at intent formation, exactly Rubicon-conformant. +— crossing at intent formation: Rubicon-conformant **by construction of the +mapping** [RULING] (the −550 ms is Libet's literature value stamped onto the +transition as the budget anchor, not a quantity measured on this system). And beneath it, the prime invariant [RULING E-NOBODY-WAITS-1]: **nobody waits for anything or any scheduling.** Writes are fire-and-forget ("melden macht @@ -190,10 +203,16 @@ envelope-auditor gate — never as ad-hoc lanes.** computing-in-superposition). The niche VSA kept is the niche the papers actually prove. 5. **The measured numbers are the vision's spine**, not its decoration: - 611M SPO lookups/sec · 1.8 ns palette distance · 17K tokens/sec · - 1–2 ms orchestration vs 8.4 s oracle · N_eff 11.00/12 · k\*=1 · - 1549 tests green at the M9 merge. When a future session doubts the - aspiration, it should re-run the probes, not re-read the prose. + N_eff 11.00/12 · k\*=1 · 1549 tests green at the M9 merge · + 1–2 ms orchestration vs 8.4 s oracle (2026-07) · and, from the + 2026-03-31 benchmark arc (dated, not re-run since): 611M SPO + lookups/sec and 1.8 ns palette distance (two distinct metrics), + 17K tokens/sec. The boldest spine claim stays named: **DeepNSM + replaces transformer inference for its scope — 680 GB model → + 16.5 MB, 50 ms/token → <10 µs/sentence** (the crate's measured + claim; its parity against the temporal-stream lane is exactly what + D-MTS-1 tests). When a future session doubts the aspiration, it + should re-run the probes, not re-read the prose. ## 8. The road (in dependency order) From 093489cb20ff3e6e67681b28ba355baaa2d566ad Mon Sep 17 00:00:00 2001 From: Claude Date: Fri, 10 Jul 2026 20:22:22 +0000 Subject: [PATCH 4/4] docs: apply PR #676 review fixes (census counts, workspace scope, MD040) coderabbit findings on the doc arc, all five verified against the tree and applied: - MODULE-TABLE ancestry census: thinking-engine heading 49 -> 51 (table rows were already correct at 51), cognitive-shader-driver heading 23 -> 22 src files, LAB-ONLY stat 11-of-23 -> 12-of-22 (the list had omitted auto_detect.rs which the table itself marks LAB-ONLY) - MODULE-TABLE scope note: p64-bridge and thinking-engine sit under workspace.exclude (Cargo.toml lines 33/36), not workspace.members - AGENT_LOG census-worker entry: post-commit metadata recorded (commit 13188ad, validation, outcome) replacing the stale 'not committed' wording; counts corrected with a dated correction note - VISION.md ancestry-pipeline fence labeled 'text' (markdownlint MD040) Co-Authored-By: Claude Claude-Session: https://claude.ai/code/session_01Awg6TXocHcwTtc6eGsHcdD --- .claude/board/AGENT_LOG.md | 4 ++-- .claude/v3/MODULE-TABLE.md | 15 +++++++++------ .claude/v3/VISION.md | 2 +- 3 files changed, 12 insertions(+), 9 deletions(-) diff --git a/.claude/board/AGENT_LOG.md b/.claude/board/AGENT_LOG.md index c82cb6e5..94e9621b 100644 --- a/.claude/board/AGENT_LOG.md +++ b/.claude/board/AGENT_LOG.md @@ -7,9 +7,9 @@ ## 2026-07-10 — census-worker-ancestry-pipeline — V3 census extended to thinking-engine / p64-bridge / cognitive-shader-driver (doc-only) -- **Census worker (Sonnet, doc-only, edit-only fleet rule honored):** appended `.claude/v3/MODULE-TABLE.md` ADDENDUM covering 49 thinking-engine files + p64-bridge/lib.rs + 23 cognitive-shader-driver src files (bare `crates/p64` does not exist as a workspace member — only `p64-bridge`). Added a `gem-status` column (WIRED-HOT-PATH / UNWIRED-GEM / CALIBRATION-ONLY / LAB-ONLY / RESIDUE) per FUTURE-DESIGN.md's wiring-queue framing. +- **Census worker (Sonnet, doc-only, edit-only fleet rule honored):** appended `.claude/v3/MODULE-TABLE.md` ADDENDUM covering 51 thinking-engine files + p64-bridge/lib.rs + 22 cognitive-shader-driver src files (bare `crates/p64` does not exist locally — only `crates/p64-bridge`, and both it and thinking-engine sit under `workspace.exclude`, not `members`). Added a `gem-status` column (WIRED-HOT-PATH / UNWIRED-GEM / CALIBRATION-ONLY / LAB-ONLY / RESIDUE) per FUTURE-DESIGN.md's wiring-queue framing. *[Correction 2026-07-10, PR #676 review: the worker's headline counts were 49/23; recount against the tree = 51/22 (the table rows were right, the headings/summary wrong); workspace-scope phrasing fixed — coderabbit findings, all applied.]* - **Notable finds:** style-table proliferation is WORSE than FUTURE-DESIGN/D-TSC-1 tracked — `superposition.rs::ThinkingStyle = DetectedStyle` collides in name (not value) with `cognitive_stack.rs::ThinkingStyle = StyleFamily` in the SAME crate; `auto_style.rs` (bare consts) + `engine_bridge.rs::UNIFIED_STYLES` + p64-bridge's `StyleParams` are 3 MORE independent 12-slot tables beyond what D-TSC-1's fleet-run addressed — worth a follow-up grep to confirm `UNIFIED_STYLES`'s definition site (not just callers) was routed. Also: `GestaltState` independently defined in both `awareness_dto.rs` and `world_model.rs`; 3 near-identical lens modules (jina/bge_m3/reranker_lens); `signed_domino.rs`/`branching.rs` extend the M8 near-duplicate-engine count beyond the 3 FUTURE-DESIGN names. `splat_ops.rs` is RESIDUE (self-declared deprecated, scheduled sprint-15+ removal, do not touch per §1 rule 8). -- **No code changed.** Working tree has the MODULE-TABLE.md addendum only; not committed. +- **No code changed.** Deliverable: MODULE-TABLE.md ADDENDUM (2). Commit: `13188ad` (landed with the VISION.md draft, same arc; count corrections in the PR #676 review-fix commit). Tests: n/a (doc-only; validation = per-file `pub`-surface grep + row-vs-tree recount). Outcome: census live, feeds FUTURE-DESIGN wiring queue + TD-STYLE-TABLE-RESIDUE extension. ## 2026-07-10 — fable-674-postmerge — #674 MERGED; E-NOBODY-WAITS-1 banked (doc-only, operator: "leave it as is and just document") diff --git a/.claude/v3/MODULE-TABLE.md b/.claude/v3/MODULE-TABLE.md index 71422d76..fbce48be 100644 --- a/.claude/v3/MODULE-TABLE.md +++ b/.claude/v3/MODULE-TABLE.md @@ -355,12 +355,15 @@ table); `nars_engine.rs` gained the runbook-keyed `style_vector_for`. > previously named. > > **Note on scope:** a bare `crates/p64` does not exist in this workspace — -> only `crates/p64-bridge` is a workspace member (`Cargo.toml` line 33). p64 +> only `crates/p64-bridge` exists locally, and both it and +> `crates/thinking-engine` are listed under `workspace.exclude` +> (`Cargo.toml` lines 33/36), NOT `workspace.members` — censused here as +> local crates, not workspace members. p64 > itself is presumably the sibling `ndarray`-hosted crate consumed by name > (`p64_bridge::CognitiveShader` docstrings reference "p64 (ndarray)" as the > topology host) — not censused here since no local crate dir exists to walk. -### thinking-engine (49 files, `crates/thinking-engine/src/*.rs`) +### thinking-engine (51 files, `crates/thinking-engine/src/*.rs`) | File | Consumes | Emits (key types) | LE contract | Function | Tech debt / duplication | gem-status | |---|---|---|---|---|---|---| @@ -438,7 +441,7 @@ p64-side topology object that `cognitive-shader-driver::driver.rs` wraps — i.e. this crate is genuinely load-bearing hot path, not a lab artifact, despite having zero tests visible in this header/pub-surface pass. -### cognitive-shader-driver (23 `src/*.rs` files + `build.rs`; tests/examples not censused) +### cognitive-shader-driver (22 `src/*.rs` files + `build.rs`; tests/examples not censused) | File | Consumes | Emits | LE contract | Function | Tech debt / duplication | gem-status | |---|---|---|---|---|---|---| @@ -479,9 +482,9 @@ having zero tests visible in this header/pub-surface pass. cognitive-shader-driver: safetensors config→codec JIT defaults) — same filename, unrelated purpose, easy to grep-confuse across crates. 3. **LAB-ONLY is the majority shape** in this crate's non-hot-path files - (codec_bridge, codec_kernel_cache, codec_research, cypher_bridge, - decode_kernel, grpc, planner_bridge, rotation_kernel, serve, token_agreement, - wire = 11 of 23 files) — consistent with the `lab-vs-canonical-surface.md` + (auto_detect, codec_bridge, codec_kernel_cache, codec_research, + cypher_bridge, decode_kernel, grpc, planner_bridge, rotation_kernel, serve, + token_agreement, wire = 12 of 22 files) — consistent with the `lab-vs-canonical-surface.md` doctrine; all self-declare LAB-ONLY in their own doc comments, so this is disciplined, not drift. 4. `mailbox_soa.rs` and `driver.rs` and `wire.rs` are the three largest files diff --git a/.claude/v3/VISION.md b/.claude/v3/VISION.md index dfc232c8..65b1c2ee 100644 --- a/.claude/v3/VISION.md +++ b/.claude/v3/VISION.md @@ -160,7 +160,7 @@ this surface is judged against the invariant's sentence first. ## 6. The body plan — the ancestry pipeline -``` +```text ladybug-rs → thinking-engine → p64 (the DTO ladder: StreamDto Φ / PerturbationDto Ψ / BusDto B / ThoughtStruct Γ) → cognitive-shader-driver → SoA (MailboxSoA, value tenants)