Wire arm-discovery (Aerial+) → self-directed reasoning: the endgame's ingestion leg

.claude/board/AGENT_LOG.md

@@ -1,3 +1,11 @@
+## 2026-07-19 — cosine-free ingestion leg: 2-agent fleet (Sonnet grindwork + Opus filigree) — main-thread orchestration
+
+- **Task:** operator ("go ahead autoattended autonomous … opus agents for filigrane, sonnet 5 agents for grindwork") → make the two comprehension findings durable Rust probes, fanned out by model tier.
+- **Fleet (edit-only, shared checkout, no worktree; orchestrator compiled centrally):** Sonnet agent → `crates/deepnsm/examples/freq_is_cosine.rs` (grindwork: the ρ test). Opus agent → `crates/deepnsm/examples/homograph_collapse.rs` (filigree: the 2³ homograph collapse).
+- **Verified centrally** (scoped `cargo clippy/run --example <name>` so the operator's in-flight `gridlake_*` WIP was NOT touched): freq_is_cosine clippy-clean, **ρ=0.762** (reproduces the Python probe exactly; KILL floor 0.5); homograph_collapse clippy-clean, thought/left/means → distinct centroids (r53/r692, r152/r4559, r130/r1392), **809/809** verb∩noun homographs collapse.
+- **Board:** `E-FREQ-IS-COSINE-REPLACEMENT-1` + `E-SURFACE-FORM-COLLAPSE-1` prepended. Commits `ea18a444` + `fb20a78e` on `claude/happy-hamilton-0azlw4` (PR #757, draft), on top of the `48cef03f` arm-discovery bridge (`E-ARM-DISCOVERY-REASONING-BRIDGE-1`).
+- **Outcome:** the cosine-free ingestion-leg arc = 3 durable probes (discover premises from data → frequency-is-cosine distance → surface-form homograph collapse → reason), all measured + KILL-gated, no core change. Fleet pattern honored (Sonnet grind / Opus filigree). Left PR in draft — operator drives ready/merge.
+
 ## 2026-07-19 — qualia-immersion / Chalmers-dissolution thesis + falsifiable exhibit — main thread
 
 - **Task:** operator ("break Chalmers hard problem by subjective experience of qualia through reasoning as immersion") → capture the endgame thesis + a falsifiable exhibit.

.claude/board/EPIPHANIES.md

@@ -1,3 +1,43 @@
+## 2026-07-19 — E-FREQ-IS-COSINE-REPLACEMENT-1 — the per-subgenre word FREQUENCY DISTRIBUTION reproduces semantic-distance ordering with NO embedding and NO cosine-on-vectors (Spearman ρ=0.762 on the 8 in-repo genres alone) — "the centroid IS the cosine", the empirical ground under the palette256 / Morton / perturbation-shader distance
+
+**Status:** FINDING (built + measured in both Python and Rust, KILL-gated ρ≥0.5). **Confidence:** High. Deliverable: `crates/deepnsm/examples/freq_is_cosine.rs` (std-only, zero-dep). Operator thread: *"the frequency translates to cosine replacement centroid"* + *"there's no way this isn't relevant for distribution"* — offered as a testable claim, tested.
+
+**The claim, tested (not asserted).** A word's per-subgenre COCA frequency vector, `ln(1+PM)` per genre, z-scored across the vocabulary, Fisher-z'd, gives a pairwise distance that reproduces the DeepNSM README's NSM semantic-distance ordering at **Spearman ρ = 0.762** — with NO embedding and NO cosine computed on any learned vector. The distance is a property of WHERE the word's frequency lands, not of a comparison. Ground truth = the 8 README NSM pairs (`i/you` closest … `say/word` farthest); the extremes land correctly, the misranks (`before/after`, `the/this`) are exactly where 8 coarse genres lose the resolution the 96-subgenre palette recovers.
+
+**Why this is the ground under the distance substrate.** It is the empirical reason the operator's architecture is "relevant for distribution": because the frequency centroid IS the cosine, (a) there is nothing to compute pairwise — distance immaterializes (the Fisher-z is baked into palette-256 level placement, not stored k×k; cf. `bgz-tensor::fisher_z::FisherZTable` as the *materialized* reference the palette is calibrated against, and `contract::distance::similarity_z` / `mean_similarity_fisher` as the transform); and (b) it amortizes over the whole gridlake through the Morton 2bit×2bit 4×4 cascade rather than O(n²) pairwise, because every word is *placed*, not *compared*. The `6× palette256:palette256` perturbation-shader tenant (3 SPO + 3 AriGraph SPO), the ARM codebook oracle (`ρ=0.9973 vs cosine`), and CAM-PQ distance are all this one immaterialized-Fisher-z object.
+
+**Honest boundary (truth-architect).** This uses only the **8 coarse genres** committed to the repo (`lemmas_5k.csv`); the full **96-subgenre** palette (gitignored) is higher-resolution, so ρ=0.762 is a **FLOOR**, not a ceiling — the README's own NSM distances came from that 96-D CAM-PQ. The claim proven is the ORDERING (rank correlation), not absolute-distance identity (the scales differ). Additive: example only, no core change. **Cross-ref:** E-SURFACE-FORM-COLLAPSE-1 (the surface→lemma-centroid front door), E-ARM-DISCOVERY-REASONING-BRIDGE-1 (the ingestion leg), `contract::distance` (`similarity_z`), `bgz-tensor::fisher_z::FisherZTable`, the operator's palette256-immaterialized-Fisher-z thread.
+
+## 2026-07-19 — E-SURFACE-FORM-COLLAPSE-1 — the SPO 2³ role mask is a HOMOGRAPH-COLLAPSE OPERATOR: a surface word form is a superposition of (lemma, PoS) readings, and fixing the SPO slot collapses it to a UNIQUE lemma → a UNIQUE lemRank (the frequency-centroid address) — same surface, different centroid by role, measured on real COCA data
+
+**Status:** FINDING (built + measured, KILL-gated). **Confidence:** High for the collapse mechanism; the S/O→nominal, P→verbal role→PoS map is an explicit simplification (see boundary). Deliverable: `crates/deepnsm/examples/homograph_collapse.rs` (std-only, zero-dep). Operator thread: *"you can hold the meaning of all words in parallel and SPO 2^3 was always of superposition S_O / SP_ / _PO"* + *"please read the meaning of the word form abbreviation to understand how it could be used."*
+
+**The claim, made runnable.** A surface token is not a symbol — it is a SUPERPOSITION of its `(lemma, PoS)` senses (`thought` = {think·v, thought·n}). The SPO 2³ role mask is the COLLAPSE OPERATOR: the grammatical slot the token fills fixes a PoS class (Subject/Object → nominal `n`, Predicate → verbal `v`), which selects exactly ONE sense → ONE lemma → ONE `lemRank`. Since `lemRank` IS the frequency-centroid address (the same 5,050-lemma space the ρ=0.762 cosine-replacement lands in), **the same surface string addresses a DIFFERENT centroid depending on its role** — homograph disambiguation as superposition collapse, no parser weights.
+
+**Measured (`word_forms.csv`, 10,239 distinct surface forms):**
+- 1,166 surfaces (11.39%) are ambiguous (>1 `(lemma,PoS)` sense); **1,164 are CROSS-PoS** (senses span ≥2 PoS letters) → collapsible by a role mask; only **2** are same-PoS residue (`found` = find·v / found·v) — the honest case a role mask ALONE cannot split.
+- **809 verb∩noun homographs, ALL 809 (100%)** collapse to DISTINCT verbal-vs-nominal centroids under P vs S/O.
+- Demonstrated (same surface, different centroid): `thought` → P **r53** (think·v) vs S/O **r692** (thought·n); `left` → P **r152** (leave·v) vs S/O **r4559** (left·n) — the `left·j` adjective sense is selected by NEITHER S/P/O mask (honest: adjective is a modifier role, not S/P/O); `means` → P **r130** vs S/O **r1392**. All three KILL-gated (each side unique, the two differ).
+
+**Why it matters for the ingestion leg.** This is the FRONT DOOR of the cosine-free ingestion (`E-ARM-DISCOVERY-REASONING-BRIDGE-1`): running text arrives inflected and ambiguous; `word_forms.csv` (11,460 forms → 5,050 lemma-centroids) resolves any surface form to its lemma-centroid, and the SPO role collapses the homograph superposition to the ONE centroid the slot selects. Surface text → role-collapsed centroid → palette/frequency distance, no embedding.
+
+**Honest boundary (truth-architect).** The S/O→nominal, P→verbal map is a SIMPLIFICATION — real syntactic slot assignment is the grammar/parser's job (a full FSM resolves which role a token fills). This probe demonstrates the COLLAPSE MECHANISM (a role mask disambiguates a homograph to a unique frequency-centroid), NOT full parsing. Additive: example only, no core change. **Cross-ref:** E-ARM-DISCOVERY-REASONING-BRIDGE-1 (the ingestion leg this feeds); the frequency-as-cosine-replacement finding (ρ=0.762); the 2³ SPO superposition (`S_O/SP_/_PO`); `deepnsm/word_frequency/word_forms.csv`.
+
+## 2026-07-19 — E-ARM-DISCOVERY-REASONING-BRIDGE-1 — the endgame's INGESTION leg wired: Aerial+ association-rule discovery mines premises from data (no LLM) → NARS-truth SPO → the self-directed loop derives the syllogism; and the memory-model comparison (RNN vs ring buffer vs Markov-NARS accretion) resolves to "the NARS revision arc IS the Markov trajectory"
+
+**Status:** FINDING (built + measured, 4 KILL-gates green) + FRAMING (memory-model comparison). **Confidence:** High for the bridge; the streaming-accretion half is plan-not-yet-wired (stated). Deliverable: `crates/lance-graph-arm-discovery/examples/discovery_to_reasoning.rs` (std-only, zero-dep, default features). Operator: *"did you wire Lance-graph-arm-discovery to arm discovery using aerial+ logic as functional routing of ideas … rnn vs ring buffer in Tesseract … does arm-discovery have any ice caking vs Markov chain context building and reasoning."*
+
+**The bridge (measured, no LLM, deterministic):** the two halves the workspace kept apart are now composed end-to-end — `extract_rules` (the REAL Aerial+ codebook probe: nearest consequent per feature, then support/confidence confirm on data) mines 4 `is_a` premises from 200 co-occurrence rows (`socrates→philosopher`, `plato→philosopher`, `philosopher→human`, `human→mortal`; each `f=1.00 c=0.98` earned from 50 co-occurrences) → `CandidateTriple::from_rule` routes them as NARS-truth SPO edges → the self-directed transitive-deduction loop (the shipped `self_directed_graph.rs` shape, std-only here) derives `socrates … mortal` (c=0.94) at a fixed point. **Nothing hand-asserted — the premises were MINED, the conclusion REASONED.** This is "read-without-an-LLM feeding reasoning-about-itself" as one float-free pipeline.
+
+**Two mining mechanics surfaced (correct the naive first cut):** (a) `extract_rules` picks the SINGLE codebook-nearest consequent per feature, so a multi-parent entity (`socrates` is_a philosopher AND greek) yields 0.5 rule-confidence and is DROPPED below the floor — a clean single-parent tree is required, and the reasoner supplies the chaining single-hop mining cannot; (b) the miner emits BOTH `f0→f1` and reversed `f1→f0` rules — the router must filter to the `is_a` (subject→object) direction. Both are in the example.
+
+**The memory-model comparison (the operator's RNN/ring-buffer/ice-cake question), grounded:**
+- **RNN** (Tesseract LSTM, transcoded byte-parity in tesseract-rs `Lstm::forward` — cell `c`, output `h`, int8-requantized into the next timestep) = unbounded history compressed LOSSILY into a fixed-dim recurrent state, baked into weights: implicit, non-inspectable, non-revisable, non-replayable "ice cake in the weights."
+- **Ring buffer** = explicit bounded FIFO, lossless-in-window, evicts the past. **arm-discovery literally contains one:** the `RatificationQueue` (bounded ring buffer, default 1024 candidates, backing onto Lance) in `streaming-arm-nars-discovery-v1.md:280`.
+- **Markov stream + NARS revision** (temporal.rs sorted stream + the self-directed loop) = explicit, unbounded, revisable, REPLAYABLE (version-range reads). **arm-discovery's cross-window ice-caking IS this:** the streaming plan accretes each rule's truth by NARS-revising against its prior each window, and `I-SUBSTRATE-MARKOV` states *"the NARS revision arc IS the Markov trajectory"* (Chapman-Kolmogorov-consistent by construction). So arm-discovery's ice-caking and Markov context-building are the SAME mechanism. The workspace's bet: replace the RNN's opaque learned recurrence with explicit NARS-revised Markov accretion — same context-building function, inspectable/deterministic/no-weights. That is "AGI without LLM" at the memory layer.
+
+**Honest boundary + next step:** the shipped example is SINGLE-WINDOW stateless (the streaming cross-window NARS-revision accretion + the ring-buffer ratification queue are the plan `streaming-arm-nars-discovery-v1`, not yet wired); the reasoner is a std-only mirror of the real `TripletGraph` loop (kept zero-dep to honor the crate's isolation). **Named follow-up (grounds "read without LLM" in real text):** replace the synthetic co-occurrence rows with DeepNSM's real parse of philosophy text — DeepNSM's 4096-rank COCA vocab (98.4% coverage; academic register carries socrates/philosopher/mortal) parses text→SPO via its 6-state PoS FSM (no LLM), whose co-occurrences ARM then mines. **Cross-ref:** E-SELF-DIRECTED-GRAPH-1 (the reasoning leg), I-SUBSTRATE-MARKOV (NARS revision = Markov), `streaming-arm-nars-discovery-v1.md` (the streaming accretion plan), tesseract-rs `E-OCR-LSTM-1` (the RNN transcode).
+
 ## 2026-07-19 — E-PERTURBATION-CONVERGENCE-1 — the five x265/H.268 arc findings CONVERGE onto ONE pinned equation (§4:108), real not rhyme; ratified by a 5-agent council; two invariants keep it from collapsing + one unproven "vertical" leg
 
 **Status:** FINDING (convergence council: 3 Sonnet harvesters [comma/phase, centroid-gridlake/256, axis-distinctness] + 2 Opus adjudicators [`convergence-architect`, `dilution-collapse-sentinel`]; both adjudicators INDEPENDENTLY ratified the same verdict — Opus for filigrane, Sonnet for grindwork per the standing model policy). Answers the operator's convergence question: *"Morton-cascade inverse-pyramid 4×4 ergonomics + centroid gridlake + vertical Pythagorean comma → perturbation-shader-cascade replayability WITHOUT materialization and WITHOUT collapsing."*