AdaWorldAPI · AdaWorldAPI · Jul 28, 2026 · Jul 28, 2026 · Jul 28, 2026
diff --git a/.claude/board/AGENT_LOG.md b/.claude/board/AGENT_LOG.md
@@ -1,3 +1,13 @@
+## 2026-07-28 — rails-shaped rung lift: 144-cell reasoning + 4×4 Morton cascade + qualia (main-thread, no subagent)
+
+- **Operator steers (4, same arc):** "use the 12×12=144 verb matrix for reasoning" (not a flat membership bit); "4×4 Morton cascade inverse pyramid perturbation shader"; "either as is or expanded to 16×16" → as-is, reserves dormant; "verbs actionability should be deterministic" → table reads only; "the qualia of blind × context would make awareness even better" → quale = cell modal × Staunen-at-lift; "(3×4)² irrational Morton drifting Pythagoras comma" → real mechanism, [S], logged as PROBE-COMMA-144 with the golden-step null as falsifier template.
+- **Built — contract:** verb_table Morton cascade layer (quadrants, morton_cell, quadrant_prior + 4 tests incl. the MEASURED residual probe: mean 0.0774 / max 0.500 Grounds.lokal, pinned); verb_lexicon IRREGULAR_PASTS + epistemic lemmas (see→Mirrors 0.70, know→Abstracts 0.85; WordNet supersense split concurs) + epistemic_reading (cue gate × matrix read) + 5 tests incl. the catalogue↔matrix weld.
+- **Built — probe:** lift gate → epistemic_reading; meta-truth discount = cell modal (never a constant); RungLift carries modal/cell/staunen_at/quale; B5 falsifier (scene-scale quale crowns reflexive 3:7, context factor must rise independently).
+- **Measured:** fixture all green (3:7 quale 0.155 ≈ 2× 3:6's 0.086); corpus-scale mean-staunen DILUTES (1:4 0.062 > 3:7 0.056) — honest finding, rate-not-count one layer up; PROBE-QUALE-LOCAL pre-registered, not tuned in.
+- **Gates:** 135 grammar tests green · clippy `-D warnings` clean ×2 · fmt clean · fixture asserts green.
+- **Codex review round (#859, both confirmed real, fixed pre-merge):** P1 — snapshot taken AFTER admit_derived leaked modal into its own context factor; fix (snapshot before output) FLIPPED the corpus verdict: blind ranking now crowns 3:7 (0.047; 1:4 = 0.000 pristine) and the earlier "mean dilutes" finding is superseded as substantially artifact. P2 — homograph irregulars (saw = tool, bore = drill) unconditionally past-classified in public classify_verb → split: AMBIGUOUS_IRREGULAR_PASTS resolve only under the cue gate (epistemic_reading); read_verb("saw") stays None (sparsity preserved), tested.
+- **Board (this commit):** EPIPHANIES E-RUNG-LIFT-RAILS-SHAPED-144-QUALIA-1 · LATEST_STATE inventory. Task #49.
+
 ## 2026-07-28 — PROBE-EYES-OPENED: the Adam awareness printed blind (main-thread, no subagent)
 
 - **Operator arc:** "the most brutal test would be to find out how the KJV bible bake needs to connect the dots about the Adam awareness — 'Their eyes were opened'" + "we could start with that and Hermeneutik" + "how does Hermeneutik compute to NARS". Built the four-blade probe over the real KJV bake (`/tmp/pg10.txt`, Gutenberg #10, local-only).

diff --git a/.claude/board/EPIPHANIES.md b/.claude/board/EPIPHANIES.md
@@ -1,3 +1,15 @@
+## 2026-07-28 — E-RUNG-LIFT-RAILS-SHAPED-144-QUALIA-1 — the rung-lift stops being a membership bit and becomes a 144-CELL READ (operator: "use the 12×12=144 verb matrix for reasoning"), the 144 gets its 4×4 Morton cascade addressing (operator: "4×4 Morton cascade inverse pyramid"; "either as is or expanded to 16×16" → as-is, reserves dormant), and awareness gains the quale = BLIND × CONTEXT (operator: cell modal × Staunen-at-lift) — which, after the codex-P1 measurement fix (snapshot BEFORE the lift's own output), crowns 3:7 in the BLIND corpus ranking at full scale.
+
+**Status:** SHIPPED + MEASURED. Contract: `verb_lexicon::{IRREGULAR_PASTS, epistemic_reading, EpistemicReading}` (epistemic lemmas minted into the matrix: SEE-class → `Mirrors` modal 0.70, KNOW-class → `Abstracts` modal 0.85 — WordNet independently makes the same cut, verb.perception vs verb.cognition, two of its 15 supersenses ≈ the 16-cell coarse grid; troponymy = the pyramid below) + `verb_table::{FamilyQuadrant, TenseQuadrant, morton_cell, same_quadrant, quadrant_prior}` (one byte per cell `[fq:2|tq:2|fm:2|tm:2]`, high nibble = coarse quadrant = nibble-ancestry, D-TILE256-shaped; 12×12 occupied in the 16×16 = palette256 page, reserves RESERVE-DON'T-RECLAIM). Probe: cell-graded meta-truth (the discount IS the cell's tense-modulated modal, never a constant) + B5 quale falsifier. Gates: 135 grammar tests green, clippy `-D warnings` ×2 clean, fixture asserts green. **Confidence:** High on all pinned numbers; the family-quadrant carve and epistemic placements are STARTER (wordnet-supersense alignment probe queued).
+
+**The rails-shaping (the operator's correction, one layer up from #842's).** #858 gated rung lifts on flat `is_perception_verb` membership — the same hand-rolling anti-pattern the archetype-consumer PR fixed for relations. Now: the cue catalogue only licenses the that-complement (a lexical fact — `resemble` is Mirrors but takes no that-clause; Moore's cheap check); the REASONING is the cell read: `(family, tense) → base_prior.combine(tense_modifier) → modal` = the lift's epistemic force, + the cell's Morton byte for carriage. The catalogue↔matrix WELD test makes drift impossible: every cue surface form must read a cell. The 144 immediately earns its keep on the first text: "the woman SAW that the tree was good" (3:6, Mirrors/Past, 0.70) < "they KNEW that they were naked" (3:7, Abstracts/Past, 0.85) — seeing and knowing graded by the matrix, Change-quadrant SIBLINGS in the pyramid (same coarse cell, different members).
+
+**The inverse-pyramid residual probe (measured, the 4×4 claim tested).** The table was secretly already pyramid-shaped — `base_prior` groups the 12 families into exactly the 4 quadrant superclasses (its own comment headers: Change/Action/State/Discovery) and `tense_modifier` is class-shaped over the 4 tense quadrants (`t/3`: Simple/Continuous/Perfect/Mood). Measured over all 144 cells × 5 axes: **mean residual 0.0774** (cells ARE small perturbations on quadrant centroids — the pyramid claim holds on the mean) with a named outlier catalogue: **Grounds.lokal 0.500** (the State class's lokal outlier), Causes.temporal 0.300, Contradicts.kausal 0.279, Mirrors.lokal 0.275 — the per-family axis signatures the coarse level deliberately does not carry. **Lokal is the axis the 4-class carve compresses worst** — a real input to the wordnet tune. Quadrant centroids keep the class signal (Action→kausal-dominant, State→modal-dominant, asserted) = graceful degradation: an unknown-but-quadrant-resolvable verb still reads class-correct priors. All deterministic (operator: "verbs actionability should be deterministic") — table reads, no stochastic scoring.
+
+**B5 — awareness = blind × context (quale = modal × Staunen-at-lift), with a codex-P1 correction that FLIPPED the corpus verdict.** First cut snapshotted the arena AFTER `admit_derived` — the lift's own modal-scaled meta-belief sat inside its own context factor, so `modal` leaked into both sides of the product and the corpus ranking crowned 1:4 (staunen 0.089, mostly contamination + early-arena entropy). Codex caught it (P1: "measure context before admitting the graded meta-belief"). Post-fix, context = the arena strictly BEFORE the lift's output: **1:4's context collapses to exactly 0.000** (a pristine, fully-confident creation arena has zero felt context — no unresolved truth, no committed contradiction), and the **blind corpus ranking over all ten lifts crowns 3:7**: quale 0.047 > 1:12's 0.041 > 3:6's 0.035 > … > 1:4's 0.000. Scene-scale (asserted): 3:6 saw = 0.70 × 0.053 = 0.037 vs 3:7 knew = 0.85 × 0.133 = **0.113** (3×), with the assert requiring the CONTEXT factor alone to rise between them (the eat reversal lands in between) so the quale is never decoration over reflexivity. The earlier "field-mean staunen dilutes at corpus scale" reading is **superseded — it was substantially the contamination artifact**; the mean-dilution mechanism remains real in principle (margins are modest: 0.047 vs 0.041), so **PROBE-QUALE-LOCAL** (Δstaunen, the shipped `Dissolution` step delta) is demoted from "needed rescue" to "optional sharpening", still pre-registered, never tuned in post hoc. Lesson for the ledger: a measurement-hygiene review fix did not just clean a number — it flipped a scientific verdict; the context factor MUST exclude the event's own output (context precedes consequence, mechanically).
+
+**Queued probes (logged, not built):** PROBE-QUALE-LOCAL (above); **PROBE-WORDNET-QUADRANTS** (families ↔ WordNet 15 verb supersenses alignment over the in-house wordnet rail, D-RCC-5-adjacent — the corpus tune for both the carve and the epistemic placements); **PROBE-COMMA-144** (operator: "(3×4)² irrational Morton cascade drifting Pythagoras comma" — REAL as mechanism: the tense carve is 3×4 and a mixed-radix (3,4) interleave never closes, 3ᵏ≢4ᵐ, the wrap defect = a comma analogue, D-QUANTGATE's coprime-walk slot; but [S] until a probe names WHAT READS THE PHASE — the golden-step null (E-BASE17-…-RELABEL-1: GOLDEN_STEP=11 carried zero information) is the falsifier template: a drifting relabel is information-free unless a reader consumes phase). **O7 fence held:** all of this addresses verb_table's 144 ONLY; the divergent sigma_rosetta 144 (E-RUNG2-TWO-144S-1) is untouched. Refs: E-EYES-OPENED-PRINTS-BLIND-1 (the probe this hardens), E-SCI-1-VERB-TABLE-ARCHETYPE-CONSUMER-AND-FSM-FEEDER-1 (the relation-side rails-shaping), persona-vs-rung-ladder.md (rung-2 = the 144 thinking atoms), task #49.
+
 ## 2026-07-28 — E-EYES-OPENED-PRINTS-BLIND-1 — PROBE-EYES-OPENED: the Adam awareness event ("their eyes were opened", Gen 3:7) PRINTS as epistemic structure, found BLIND on real KJV Genesis 1–4 (106 verses) — across TEN rung lifts exactly ONE is self-referential and it is 3:7 `they —knew→ naked`; the blind contradiction ranking surfaces the fall's two hinges (`eat`, `die`, both 0.850) PLUS two genuine reversals nobody designed for (`god→good` @2:18, `they→respect` @4:4-5); and the hermeneutic circle gets its missing TERMINATION PROOF from NARS.
 
 **Status:** SHIPPED + MEASURED. `crates/lance-graph-planner/examples/probe_eyes_opened.rs` (self-asserting 13-verse KJV fixture in CI; real-corpus mode `-- /tmp/pg10.txt` local-only) + `contract/grammar/clause_cues.rs` gains `is_negation` / `is_perception_verb` (exact catalogues, fire+silent tests). Gates: contract 6 unit + 4 doctests green, clippy `-D warnings` clean ×2, fmt clean, fixture falsifier asserts green. **Confidence:** High on every asserted number (fixture pinned to measurement); the extractor is deliberately COARSE (single-scene pronoun coreference → one `they` referent; no conditionals; no object scoping) and the real-corpus run is informational, not asserted.

diff --git a/.claude/board/LATEST_STATE.md b/.claude/board/LATEST_STATE.md
@@ -1,3 +1,10 @@
+## 2026-07-28 — branch `claude/x265-x266-plans-review-h9osnl` — rails-shaped rung lift (144-cell reasoning + Morton cascade + qualia)
+
+### Current Contract Inventory — verb matrix cascade + epistemic reading
+- `lance_graph_contract::grammar::verb_table::{FamilyQuadrant, TenseQuadrant, morton_cell, same_quadrant, quadrant_prior}` (`E-RUNG-LIFT-RAILS-SHAPED-144-QUALIA-1`) — the 4×4 Morton cascade addressing of the 144: one byte per cell `[fq:2|tq:2|fm:2|tm:2]`, high nibble = coarse quadrant (nibble-ancestry, D-TILE256-shaped), 12×12 occupied in the 16×16 palette256 page, reserves RESERVE-DON'T-RECLAIM. Inverse-pyramid residual probe MEASURED + pinned: mean 0.0774, max 0.500 = Grounds.lokal (outlier catalogue named; lokal = the axis the carve compresses worst). Deterministic throughout.
+- `lance_graph_contract::grammar::verb_lexicon::{IRREGULAR_PASTS, EpistemicReading, epistemic_reading}` — the rails-shaped rung-lift condition: cue gate (that-complement licensing) + 144-cell read (tense-modulated modal = epistemic force + Morton address). Epistemic lemmas minted: SEE-class → Mirrors (0.70), KNOW-class → Abstracts (0.85) — WordNet's verb.perception/verb.cognition supersense split independently confirms the cut. Catalogue↔matrix WELD test forbids drift. Irregular pasts (knew/saw/understood/…) classify with correct tense.
+- `probe_eyes_opened` B5 — awareness quale = blind × context (cell modal × Staunen-at-lift), context snapshotted BEFORE the lift's own output (codex P1 fix — the first cut leaked modal into both factors and the fix FLIPPED the corpus verdict): scene-scale asserted (3:7 = 0.85×0.133 = 0.113 > 3:6 = 0.70×0.053 = 0.037, both factors independently); corpus-scale the BLIND ranking over ten lifts crowns 3:7 (0.047; 1:4's pristine arena reads exactly 0.000 context). Codex P2 also fixed: homograph irregulars (saw/bore) stay sparse in the public classifier, resolving only under the cue gate. PROBE-QUALE-LOCAL demoted to optional sharpening. Also queued: PROBE-WORDNET-QUADRANTS, PROBE-COMMA-144 ([S] until a phase-reader is named). O7 fence held. Detail: EPIPHANIES + AGENT_LOG 2026-07-28.
+
 ## 2026-07-28 — branch `claude/x265-x266-plans-review-h9osnl` — PROBE-EYES-OPENED (the Adam awareness printed blind from the KJV bake)
 
 ### Current Contract Inventory — new clause_cues catalogues + probe example