feat(causal-edge): hardwire NAL syllogism figure resolution on CausalEdge64 (the capstone)

.claude/board/AGENT_LOG.md

@@ -1,3 +1,27 @@
+## [Main thread / Opus, autoattended] 5-agent council reviewed the NAL syllogism capstone → SOUND kernel, integration roadmap spec'd
+
+**Branch:** claude/jolly-cori-clnf9. Council R1–R5 (Opus, full-file reads, E-READ-NOT-GREP). **R1** NAL-correct (figures/rules/premise-orders byte-match canonical OpenNARS; omitted Comparison/Analogy intentional — need `<->` copula CE64 can't carry). **R2** firewall+layer SOUND; the 3 truth-fns verified byte-identical to BOTH ndarray::hpc::nars AND forward(). **R3** SOUND + applied doc fix (predicate is a *typed placeholder*; compose_p won't synthesize the induced/abduced relation). **R4** SOUND + applied doc fix (EW64 fold must be **slot-0-anchored**, not a blind left-fold that None-cascades; EdgeRef→CE64 must honor family via OGIT class + 1-based local). **R5** FIX-NEEDED *at integration only* — kernel sound, but zero callers; the capstone (3-path glue / rung elevation / cranelift↔elixir) is unwired.
+
+**Applied:** R3+R4 doc tightenings + 1 redundant-link + fmt of test code (syllogism.rs now 14 tests / clippy / fmt / rustdoc all clean). **Spec'd** the roadmap → `.claude/specs/nal-syllogism-capstone-v1.md`. Highest-leverage next step (R5 #3): promote `notation()` → `const FIGURE_RULES` table + dual `jit_template()` / `elixir_clause()` emitters — the literal "NAL notation and Elixir complete each other" (one table, two backends, offline codegen). Next: PR + subscribe.
+
+---
+
+## [Main thread / Opus, autoattended] NAL syllogism FIGURE resolution hardwired on CausalEdge64 (the capstone)
+
+**Branch:** claude/jolly-cori-clnf9. **Tests:** `causal-edge` syllogism 14 green (v2 default) / 13 (v1, the mantissa test gated); new file clippy- + fmt-clean (the 15 pre-existing `edge.rs` -D-warnings + fmt diffs are the documented v1/v2 mantissa minefield — untouched). User steered: "hardwire syllogism resolution like SPO 2³ … using causaledge64, wiring EW64"; "NAL notation = missing capstone glueing 3 reasoning methods + 10-rung ladder + JITson/cranelift vs elixir"; "34+ opennars vocabulary just needs wiring."
+
+**Did:** new `causal-edge::syllogism` — `Figure{Chain,ChainRev,SharedSubject,SharedObject}` resolved by integer SPO-palette term-sharing (the Pearl-2³ analogue); `CausalEdge64::figure()`/`syllogize()` emit the conclusion edge (outer terms + canonical NARS truth + signed mantissa + AND mask). Grounded by full reads (E-READ-NOT-GREP): nars_engine, cognitive_codebook, ndarray::hpc::nars, atoms, cognitive_shader, episodic_edges, causal-edge edge/tables. **Reverted** the speculative 3rd-copy syllogisms in `contract::exploration::NarsTruth` (mislabeled ind⇄abd vs canonical). Next: PR + 5-agent council review; then (gated) EW64→CE64 wiring in the driver.
+
+---
+
+## [Main thread / Opus] 5-agent RESEARCH council — 8 semantics/embedding papers, firewall-filtered
+
+User: "use research council 5 agents [on 8 PDFs]; grep/sed/tail/head fragments forbidden; test the reading tools first." **Test caught a blocker:** the Read tool's PDF path needs poppler (absent) — every agent would have failed. Fixed: extracted PDFs→full-text `.txt` via pymupdf (pip), recovered an 8th that was a saved MHTML web page; verified Read works on `.txt`. THEN dispatched A1–A5 (Opus, read-only, **READ full text, never grep/head/tail**). All 5 returned full-read verdicts.
+
+**Result** (doc `research-council-semantics-papers-2026-06.md`, `E-RESEARCH-COUNCIL-PROPOSE-VALIDATE`): firewall cleanly filtered 8 LLM/float papers → 1 ADOPT-NOW (SemDiD→`head2head::WinnerCriterion::Repulsion`, cosine→Hamming), 3 integer validators (Legality / `⟨u,v⟩` / footprint), 1 shared novelty-gated-selection operator (head2head + EW64), 1 adversarial foundation-probe (Kozlowski: does hard-basin CAM-PQ discard entangled low-rank semantics?), 1 SKIP (segmentation = trap). The "PROPOSE/ADDRESS" doctrine was independently re-derived by 4/5 reviewers from 7 papers. Next: @jan picks the first build (SemDiD-adopt vs Kozlowski-probe vs Legality-validator).
+
+---
+
 ## [Main thread / Opus, autoattended] 5-agent dev council → D-ATOM-4/RawEdge shipped (① Heel-compose REFUTED, ② RawEdge built, ③ deferred)
 
 **Branch:** claude/jolly-cori-clnf9. **Cargo:** contract lib **550 green** (+5 counterfactual); default clippy clean. User: "use the 5 agent development council." Convened R1–R5 (Opus, read-only); consolidated + auto-resolved + built.

.claude/board/EPIPHANIES.md

@@ -1,3 +1,38 @@
+## 2026-06-01 — E-NARS-FIGURE-CAPSTONE — NAL syllogism resolution hardwired on CausalEdge64 like Pearl 2³: the figure = which SPO palette term two edges share → rule → conclusion edge
+
+**Status:** SHIPPED (`causal-edge::syllogism`, branch claude/jolly-cori-clnf9; 14 tests v2 / 13 v1, the new file clippy- + fmt-clean). User: "the syllogism resolution needs to be hardwired similar to SPO 2^3 rung decomposition … using causaledge64 and wiring EW64" + "NAL syllogism notation is the missing capstone for glueing all 3 reasoning methods with the 10-rung ladder and the JITson cranelift templates vs elixir."
+
+**The capstone:** `CausalEdge64::forward()` composes two edges POSITIONALLY with a *pre-set* inference type — it never asks WHICH TERM they share. NAL syllogism IS that question. `figure(other)` resolves it by integer SPO-palette equality — the hardwired analogue of the Pearl 2³ mask (O(1), branch-minimal, no float on the structural path):
+- `o1==s2` → Chain → **Deduction** ⊢ s1→o2
+- `s1==o2` → ChainRev → **Deduction** ⊢ s2→o1
+- `s1==s2` → SharedSubject → **Induction** ⊢ o1→o2
+- `o1==o2` → SharedObject → **Abduction** ⊢ s1→s2
+- same statement (`s1==s2 ∧ o1==o2`) → None (that is Revision, not a syllogism).
+`syllogize(other)` emits the conclusion `CausalEdge64` (outer terms + canonical NARS truth + signed v2 mantissa Ded +1 / Ind +2 / Abd −1 + AND-ed Pearl mask). Firewall: integer term-match PROPOSES the figure; the deterministic NARS truth-function ADDRESSES. Truth math is byte-identical to `ndarray::hpc::nars` (hardware) + `forward()` (protocol) — NOT a new truth type.
+
+**The "don't reinvent" catch (E-READ-NOT-GREP in action):** user flagged "we have 34+ opennars vocabulary, it just needs to be wired." Reading-first found it all already present — vocabulary (`cognitive_codebook::{NarsInference 10, NarsCopula 12}` with the `{M-->P,S-->M}` notation in comments + `CognitiveAddress` + fingerprints), the tested engine (`planner::nars_engine::nars_infer`, 9 rules / 17 tests), the canonical truth (`ndarray::hpc::nars::NarsTruth` + 7 truth-fns), the atoms (`atoms.rs` Operation lane: abduct/deduce/induce/synthesize), the wire (`causal_edge::InferenceType` signed mantissa). The speculative deduction/induction/abduction I had started adding to `contract::exploration::NarsTruth` were a 3rd copy AND mislabeled (induction⇄abduction swapped vs the canonical engine) — **REVERTED.** The gap was never the truth math; it was the FIGURE decomposition. Next (gated): wire EW64 `EdgeRef`→`CausalEdge64`→`syllogize` across the ≤4 hot edges in cognitive-shader-driver; cranelift/elixir dual-compile of the figure table.
+
+Cross-ref: `E-READ-NOT-GREP`, the firewall doctrine, CausalEdge64 v2 mantissa, `episodic_edges` (EW64), `atoms` Operation family, the Pearl 2³ mask.
+
+---
+
+## 2026-06-01 — E-RESEARCH-COUNCIL-PROPOSE-VALIDATE — 8 LLM/float semantics papers, council-firewall-filtered: the "PROPOSE (float, offline) / ADDRESS (integer, hot)" doctrine independently re-derived from 7 of 8; 1 ADOPT-NOW, 3 integer validators, 1 adversarial foundation-probe
+
+**Status:** FINDING (5-agent research council, 2026-06-01; ALL papers READ IN FULL per `E-READ-NOT-GREP` — A1 read 1311 lines / A5 read 1517+1695, via Read not grep). Doc: `research-council-semantics-papers-2026-06.md`.
+
+**The corroboration:** all 8 papers are LLM/float-based; the council extracted only the deterministic/integer/offline kernel from 7 and cleanly SKIP'd the 1 with no quarantine seam (segmentation = unconditional firewall trap, float-all-the-way-down). **The workspace's "similarity PROPOSES (float, offline, upstream) / CAM ADDRESSES (integer, hot, deterministic)" doctrine was independently re-derived by 4 of 5 reviewers from 7 different papers** — strong external evidence the core architecture is right.
+
+**The slate (firewall-filtered):**
+- **ADOPT-NOW (offline-buildable):** SemDiD (2506.23601) → a `head2head::WinnerCriterion::Repulsion` — repulsion-from-nearest-rival (VSA-overlap, cosine→Hamming) + ε-quality-floor + harmonic combiner; training-free, integer, self-contained in `contract::head2head`. *(A4)*
+- **Shared operator (A4):** "retain a candidate iff its hypervector is far from incumbents under a quality floor" — mount on head2head (arbitration, SemDiD) AND EW64 admission (LaMAR's "novelty beats volume" — refines the just-merged `promote`/`coldest`).
+- **3 integer VALIDATORS:** Legality Score (A1, prime-reduction purity, deepnsm-side, offline) · `⟨u,v⟩` cognitive-load (A2, CAM-PQ similarity self-explanation) · footprint `{→,←,∥,#}` (A5, aerial→DOLCE ordering validator).
+- **ADVERSARIAL PROBE (highest-value foundation test):** Kozlowski (2508.10003) — non-orthogonality is *signal*; the hard 4096-basin + ρ=0.9973 CAM-PQ partition **may discard the entangled low-rank semantics**. Distinction (A3): role-disjointness (binding) is fine; *content* orthogonality is the challenge. Runnable test: antonym-direction interference vs CAM-PQ mis-addressing; pair with SAFARI's Weyl Semantic-Shift auditor over OGIT/DOLCE (tests #444 98.6%).
+- **SKIP:** segmentation (2412.08671) — firewall trap.
+
+**Process:** the council read full text (no grep/head/tail) per `E-READ-NOT-GREP`; the reading-path TEST caught a missing-poppler blocker before dispatch (extracted PDFs→.txt via pymupdf). Cross-ref: `E-READ-NOT-GREP`, the firewall doctrine, `head2head`, `episodic_edges`, DeepNSM.
+
+---
+
 ## 2026-06-01 — E-READ-NOT-GREP — judgment-critical review agents must READ full files, not grep/sed/head/tail; fragments invalidate judgment
 
 **Status:** IRON RULE (process; user-stated 2026-06-01). Across the EW64 council + the 3 prior relayed sessions, every wrong framing came from narrating off grep/head fragments rather than reading the type. The council's R3 even found a grep-induced mis-citation (`edge.rs:750 concern_level` reads `direction()`, not `PlasticityState`). **RULE:** when a review/council/grounding agent's VERDICT depends on a type's semantics, its brief MUST instruct it to READ the relevant files in full (the Read tool), NOT grep/sed/head/tail. A fragment seen out of context produces a confident-but-wrong judgment — **grep is for LOCATING, reading is for JUDGING.** Baked into the agent-brief template (`autoattended-multiagent-pattern.md` §Rule 7). Cross-ref: `E-BASIN-NOT-EDGE-PLASTICITY` (the conflation grep-fragments produced).

.claude/knowledge/research-council-semantics-papers-2026-06.md

@@ -0,0 +1,41 @@
+# Research council — 8 semantics/embedding papers, firewall-filtered (2026-06-01)
+
+**READ BY:** integration-lead, truth-architect; anyone mining external research into the substrate.
+**Status:** SYNTHESIS (5-agent research council, all papers read IN FULL — no grep/head/tail, per `E-READ-NOT-GREP`). Each verdict is firewall-checked against the deterministic CAM-PQ substrate.
+**Method:** 5 Opus reviewers (A1 NSM / A2 interpretable-embeddings / A3 embedding-structure / A4 sequence+decoding / A5 applications+firewall-critic) each read their assigned full-text papers and judged: mechanism, workspace relevance, firewall (does adoption pull LLM/float/language onto the hot path?), verdict.
+
+## The slate (8 papers)
+
+| # | paper (arXiv) | the extractable mechanism | verdict | firewall |
+|---|---|---|---|---|
+| 1 | 2505.11764 **Towards Universal Semantics / DeepNSM** (OSU) | **Legality Score** `α·(primes−molecules)/words` + lemma-containment circularity; prime-vs-**molecule** ontology | PROBE | gate **PASS** (integer, no LLM); the LLM machinery **CONFLICT** |
+| 2 | 2410.03435 **Interpretable Text Embeddings (CQG-MBQA)** | **`cognitive load = ⟨u,v⟩`** (count of shared named axes) — self-explaining similarity; contrastive cluster-vs-hard-negative axis design (offline) | PROBE | metric **PASS** (integer hot-path); learned encoder **CONFLICT** |
+| 3 | 2512.00852 **One Swallow / SAFARI** (NUS) | **Semantic-Shift basin-boundary detector** w/ Weyl-1912 bound `‖A_y‖₂·σ_max(A_x)` (no SVD, no labels) | PROBE | **PASS** (offline auditor; spectral-norm on ±1 fingerprints) |
+| 4 | 2508.10003 **Semantic Structure in LLM Embeddings** (Kozlowski, U Chicago) | **cosine-proportional interference law**; meaning = continuous ~3-axis (Eval/Potency/Activity) low-rank manifold; **non-orthogonality is signal** | **PROBE (adversarial red-team)** | N-A as component; **CONFLICT as worldview** |
+| 5 | 2510.18046 **LaMAR** (sequential recommenders) | **novelty-gated admission** — a near-duplicate signal *hurts*; diversity carries the gain | PROBE | abstracted (offline-frozen symbol + novelty gate) **PASS**; LLM gen **CONFLICT** |
+| 6 | 2506.23601 **SemDiD** (semantic-guided diverse decoding) | **repulsion-from-nearest-rival + ε-quality-floor + harmonic (weakest-link) combiner**, k parallel arcs | **ADOPT-NOW** (cosine→Hamming) | **PASS** (training-free; swap float-cosine for VSA-overlap) |
+| 7 | 2504.21074 **LLMs for Process Mining** | **footprint fitness** `(A×A)→{→,←,∥,#}`, quality = matching-cell fraction (frequency-free, integer) | PROBE (validator ADOPT-grade) | **SAFE iff quarantined** as proposer; TRAP if LLM-edges land as fact |
+| 8 | 2412.08671 **Deep Semantic Segmentation** | neighbor-window (3×3) offset refinement | **SKIP** | **TRAP** — float/learned all the way down, no propose/validate seam |
+
+## The three unifying threads
+
+1. **The firewall is a *useful filter*, externally corroborated.** All 8 papers are LLM/float-based; the council extracted only the deterministic/integer/offline kernel from 7 and cleanly SKIP'd the 1 with no quarantine seam. **The "similarity PROPOSES (float, offline, upstream) / CAM ADDRESSES (integer, hot, deterministic)" doctrine was independently re-derived by 4 of 5 reviewers from 7 different papers** (A1 LLM-proposes-explication / Legality-validates; A2 LLM-authors-axes / ⟨u,v⟩-explains; A4 generate-arcs / repulsion-selects; A5 ARM-or-LLM-proposes-edges / footprint-validates). Strong outside evidence the core architecture is right.
+
+2. **One shared operator across the memory + inference axes (A4):** *"retain a candidate iff its hypervector is far enough from the current incumbents, under a quality floor."* Mount once on **head2head** (arc arbitration — SemDiD) and once on **EW64** (episodic-edge admission — LaMAR's "novelty beats volume"). Integer VSA-overlap vs an incumbent set; no learned weights, no float, no language. The EW64 side refines the just-merged `promote`/`coldest` (#447/#448): admission should gate on **novelty**, not recency alone.
+
+3. **Three deterministic integer VALIDATORS to adopt** (each turns a proposal into a check): **Legality** (A1, prime-reduction purity, DeepNSM-side), **⟨u,v⟩** (A2, similarity self-explanation, CAM-PQ-side), **footprint `{→,←,∥,#}`** (A5, ordering-relation validation, aerial→DOLCE boundary).
+
+## The adversarial finding (highest-value to weigh)
+
+**A3/Kozlowski (#4) is the strongest external challenge to the spine geometry.** Cross-model-replicated: semantic features are **meaningfully non-orthogonal** (whitening *reduces* human-rating fidelity ~20% — "non-orthogonality is a feature, not a bug"), and meaning collapses to a continuous ~3-axis low-rank manifold. A3's key distinction: **role disjointness** (SUBJECT vs PREDICATE binding slots) is standard, correct VSA — *not* challenged. But **content orthogonality** is: if the substrate treats concept-overlap as quantization noise, this says that overlap *is* the carrier of meaning, and the hard 4096-basin + ρ=0.9973 CAM-PQ partition **may be discarding the entangled low-rank structure that is the actual semantics.** A3's complement (#3 SAFARI) hands the *corroborating* tool: the Weyl Semantic-Shift auditor directly tests the #444 98.6%-locality (sharp `µ+3τ` spikes at class boundaries = clean partition; smeared = falsified).
+
+## Prioritized build/probe slate (firewall-filtered)
+
+- **ADOPT-NOW (offline-buildable):** **SemDiD → a head2head `WinnerCriterion::Repulsion`** — repulsion-from-nearest-rival (VSA-overlap) + ε-quality-floor + harmonic combiner. Self-contained in `contract::head2head`; integer; offline-testable. *(A4)*
+- **ADOPT-grade VALIDATOR, offline:** **Legality Score** in `deepnsm` — prime/molecule counting + circularity; a float-free certifier of SPO prime-reduction. *(A1)*
+- **PROBE (highest-value VALIDATION of the foundation):** the **Kozlowski antonym-direction interference test** — build antonym-difference directions in VSA space; does cross-direction cosine predict CAM-PQ mis-addressing? If yes, the basin model is lossy w.r.t. gradient semantics. Pair with the **SAFARI Weyl basin-coherence auditor** over the OGIT/DOLCE partition (tests #444). *(A3 — both)*
+- **PROBE (refines shipped work):** **EW64 novelty-gated admission** — gate `promote` on dissimilarity-to-incumbents, not recency alone. Needs edge-fingerprint access (the co-addressed content), so it rides the gated SoA/hot-path seam, not pure-contract. *(A4/LaMAR)*
+- **PROBE (interpretability):** **⟨u,v⟩ self-explanation** for CAM-PQ (list shared codebook IDs/primes behind a match) *(A2)*; **footprint validator** at the aerial→DOLCE landing *(A5)*.
+- **SKIP:** segmentation (#8) — unconditional firewall trap.
+
+*Cross-ref:* EPIPHANIES `E-RESEARCH-COUNCIL-PROPOSE-VALIDATE`; `E-READ-NOT-GREP` (the council read full text); the firewall doctrine (`E-ENGLISH-BIFURCATES`, markov_soa SoC); `head2head`, `episodic_edges`, DeepNSM. Papers staged at `/tmp/papers/*.txt`.