D-DIA-V2-A: insight/mush S10 detector — null falsifier caught + fixed 2 formula confounds

.claude/board/EPIPHANIES.md

@@ -1,3 +1,13 @@
+## 2026-07-23 — E-S10-COHERENCE-CLOSURE-DENSITY-1 — the S10 insight/mush null falsifier (D-DIA-V2-A) did its job on the FIRST build: it scored real=null=0 under the plan's draft formula and forced a two-part correction, converting a plausible-but-broken detector into a working one. The plan's S10 was `insight = clamp(Δcoherence + Δwonder − Δentropy, 0, 1) · [yield>θ]` with `coherence = (derived/total)·mean_expectation(derived)`. The mandatory size-preserving null (`insight_beats_size_preserving_null`: three deep `is_a` chains vs a SplitMix64 same-edge-count random rewiring) scored **real insight 0, null insight 0** — zero discrimination. The falsifier isolated TWO independent confounds, both fixed
+
+**Status:** FINDING (measured: the null falsifier now passes real 0.82 > null; 3 insight tests + 46 nars-module tests green, clippy -D warnings clean). **Confidence:** High on the correction; the detector is now a WORKING discriminator on the synthetic chain-vs-random control, not yet exercised on a whole-book step (that is the V2-A→whole-book wire, next).
+
+**Confound 1 — the `·mean_exp` coherence multiplier (deduction-attenuation inversion).** NAL deduction *correctly* attenuates confidence with chain depth (`c = c₁c₂f₁f₂`), so the DEEPEST, most-coherent transitive chains earn the LOWEST mean expectation. Multiplying closure by `mean_exp` therefore dragged coherence DOWN exactly when the graph closed MOST — the metric fought the very structure it was meant to reward. **Fix:** `coherence = closure density = derived/total` — a size-invariant ratio (coherent chains close densely: 30 edges → 135 derived → 0.82; random rewiring closes sparsely). This is the size-invariant structural signal S10 must key on.
+
+**Confound 2 — the `−Δentropy` insight term (a VSA-world import that doesn't transfer to term-logic).** `truth_entropy` = confidence-distribution Shannon spread. On EVERY productive term-logic step it RISES (premises peak at one confidence bin; deduction spreads derivations across all bins) — measured `Δent = +0.88` on the coherent step, swamping `Δcoh = +0.82` → clamp to 0. The free-energy "entropy descends on insight" pole is real in the **VSA-codebook world** (where `likelihood = cosine-vs-codebook` and a good match sharpens the distribution), but it does NOT transfer to the pure **term-logic arena**: there is no codebook, and any confidence-based surprise measure *anti-correlates* with coherent depth for a CORRECT epistemic reason (deep transitive conclusions genuinely ARE weakly held — "Abraham's father's…'s father is a man" is uncertain after many hops). **Fix:** `−Δent` is REMOVED from insight (`insight = clamp(Δcoh + Δwonder, 0, 1)·gate`); entropy stays in the mush **stall** term (`stall = entropy·(1−|Δcoh|)`) where its sign is CORRECT — entropy WITH a coherence change is attenuated (low stall), entropy WITHOUT coherence change IS the stall pole of mush. The asymmetry is the resolution: entropy-paired-with-coherence = insight; entropy-alone = mush.
+
+**The lesson banked.** (1) The size-preserving null is not ceremony — it caught a detector that scored identically on structured and random input, the exact D-SRS-3b "composite = size" failure shape, and forced the fix before the code shipped. (2) **Free-energy poles are substrate-specific:** a term borrowed from the VSA-codebook free-energy compose (`(1−likelihood)+kl`) can invert sign in the term-logic arena; port the CONCEPT (surprise descends on insight), re-derive the PROXY per substrate, and let the null falsifier arbitrate. Refs: `E-BASIN-WIDTH` (the size-preserving-null discipline), `E-WHOLE-BOOK-REASONING-RUN-1` (the arena this scores steps over), `nars/insight.rs` (`Snapshot`/`InsightMush`/`detect`/`flow_state`), `dialectic-engine-v1.md` §1 S10, `dialectic-engine-v1.md` §4 (D-DIA-V2-A).
+
 ## 2026-07-23 — E-WHOLE-BOOK-REASONING-RUN-1 — the whole KJV ran through the ACTUAL reasoning layer (not the encode) for the first time, via the SoC seam, and it falsified two real things while validating the core. `deepnsm-v2 bible_wave --export` (inbound leg) emitted the whole-book belief stream (31,327 SPO → 21,714 distinct statements: 3,072 `is_a`/copular rows, 28,255 verb rows); `lance-graph-planner --example reason_whole_book` (reasoning layer) ingested it into a `BeliefArena` and ran copula-gated `close_transitive` + the five tactics. This is the first time the dialectic engine reasoned over a whole real corpus
 
 **Status:** FINDING (measured, reproducible: the two examples + `/tmp/pg10.txt` PG#10 KJV, local-only). **Confidence:** High on the measured results; the concept-lift and Morton-scale directions are next-work, not built.

.claude/board/STATUS_BOARD.md

@@ -6,7 +6,8 @@ Plan: `.claude/plans/dialectic-engine-v1.md` (six operator pillars + S1-S12 synt
 |---|---|---|---|---|
 | D-DIA-V0 | Belief arena falsifying slice (triple-keyed + in-place stamped revision + copula-gated transitivity) | lance-graph | Shipped (all 4 registered gates green first run) | `deepnsm-v2/src/belief.rs`; 90 tests + clippy clean; `E-DIALECTIC-ENGINE-SYNTHESIS-1` |
 | D-DIA-V1 | The five tactics (RCR/TR/ASC/CAS/CR) over the Belief arena + throttles + ReasoningGap | lance-graph | Shipped (PR #816; lance-graph-planner `nars/{belief,tactics}` over `TruthValue`; 17 nars + 233 planner tests) | plan §3, §4 |
-| D-DIA-V2 | The loop: bias→recipe LUT, byte-lane council, texture window, insight/mush + null falsifier | lance-graph | Queued | plan §4 |
+| D-DIA-V2-A | insight/mush S10 detector + size-preserving null falsifier | lance-graph | Shipped (`nars/insight.rs`; null falsifier caught + fixed 2 formula confounds — `E-S10-COHERENCE-CLOSURE-DENSITY-1`; 3 insight + 46 nars tests, clippy clean) | plan §1 S10, §4 |
+| D-DIA-V2-B | The loop: bias→recipe tactic-LUT, byte-lane council, texture window | lance-graph | Queued | plan §4 |
 | D-DIA-V3 | Dissolution → field rung-elevation (the cathedral floors) + Staunen↔Wisdom flow accounting | lance-graph | Queued | plan §4 |
 | D-DIA-V4 | 64k SIMT lowering (Boolean-reachability semiring + second-pass truth) | lance-graph | Queued | plan §4, S1 |
 | D-DIA-V5 | Reach-out felt integration (dull shadow vs insight) + qualia ablation falsifier | lance-graph | Queued | plan §3.6, S12 |

.claude/plans/dialectic-engine-v1.md

@@ -179,9 +179,49 @@ stamps) and felt form (Datapath, texture) are one event read from two buckets.
   > exploration), `FlowState`/`mul`, `kanban` (contract + planner
   > style_strategy) before wiring. Reuse the one engine; probe-first on any
   > "detects insight" claim.
+  > **✅ V2-A SHIPPED (2026-07-23):** `nars/insight.rs` — `Snapshot`/
+  > `InsightMush`/`detect`/`flow_state`, reusing contract `GraphSignals` +
+  > `FlowState` (nothing invented). The MANDATORY null falsifier
+  > (`insight_beats_size_preserving_null`) did its job on the first build:
+  > it scored real=null=0 under the draft `clamp(Δcoh+Δwonder−Δent)` formula
+  > and forced a TWO-part correction (`E-S10-COHERENCE-CLOSURE-DENSITY-1`):
+  > (1) `coherence = closure density (derived/total)` — the `·mean_exp`
+  > multiplier inverted under NAL deduction attenuation (deep chains earn the
+  > lowest expectation); (2) `−Δentropy` REMOVED from insight (confidence-
+  > spread rises on every productive term-logic step — a VSA-codebook pole
+  > that does not transfer to term-logic); entropy's correct home is the mush
+  > `stall` term. Final: `insight = clamp(Δcoh+Δwonder,0,1)·[yield>θ]`. 3
+  > insight + 46 nars tests green, clippy clean. NOT yet wired to a whole-book
+  > step (V2-A→whole-book measurement is next). THEN V2-B below.
 - **V3:** dissolution detection + field elevation (S11) — the cathedral floors;
   Staunen↔Wisdom flow accounting; epiphany attractors (rate-normalized, S9).
 - **V4:** the 64k SIMT lowering — Boolean-reachability semiring + truth second
   pass (S1), masks, sweeps — only after V0–V3 green at small scale.
+  > **Column size is a capacity knob, NOT a cache constant (operator, 2026-07-23).**
+  > The 64k column (64k×512 B = 32 MB) is a cache convenience (server-L3-resident
+  > working set), not an architectural constant — 256k/128 MB or 512k/256 MB are
+  > easily affordable. The knob to grow is column CAPACITY (rows in RAM); the
+  > INVARIANT knob is the Morton TILE (the swept, cache-resident unit). They
+  > DECOUPLE under Morton-tile top-k: you sweep one tile, never the whole column,
+  > so cache behavior is invariant to total column size. Morton width scales fine
+  > (64k axis = u16 → u32 code; 512k = u19 → u38, still u64). What growing the
+  > column DOES change: brute O(N²) pair enumeration goes 16×/64× worse
+  > (`close_transitive`'s book-scale 92k-derived / 12–17 s already shows the
+  > shape) — so a bigger column makes the Morton-tile top-k substrate MANDATORY,
+  > not optional. Column growth is affordable ONLY with the retrieval mechanism,
+  > which is exactly why V4 is the Morton lowering, not a wider brute sweep.
+  > (Distinct from the GUID cascade's per-tier 64k = 256×256 centroid tile, which
+  > is codebook cardinality / canon — untouched by column length.)
+  > **LAB ceiling = 4M rows / 2 GB (operator, 2026-07-23).** The three regimes:
+  > (1) **production** = 64k / 32 MB (L3-cache-resident, the hot canonical size);
+  > (2) **affordable growth** = 256k–512k / 128–256 MB (DRAM, still cheap);
+  > (3) **LAB PoC ceiling = 4M / 2 GB** (`4,194,304 × 512 B = 2 GiB` exactly) —
+  > the upper bound for an *exceptional* proof-of-concept, **correctness-first,
+  > optimize later** (the lab-vs-canonical posture applied to field size). A 4M
+  > field is fine to HOLD resident and prove a result over; it is NOT fine to
+  > brute-sweep (O(N²) = 1.6×10¹³ pairs), so even a lab PoC at 4M runs the sweep
+  > Morton-tiled — the ceiling raises CAPACITY, never licenses brute enumeration.
+  > Morton width still fits: 4M axis = u22 → u44 code (u64). The lab result is a
+  > falsifier; the Morton-tile top-k is the production optimization that follows.
 - **V5:** reach-out integration (spider/arXiv → §3.6 felt criterion) + the
   qualia ablation falsifier (S12).