D-HOUSE-1: PROBE-HOUSE-DIFFERENTIAL-1 with the periphery as parallel strata — variant 1 KILL; variant 2 base path PASS, council-gated S3 arm inconclusive

.claude/board/EPIPHANIES.md

@@ -51,6 +51,26 @@ Theorem 2 is NOT the lattice cutoff.
 downstream of it are structurally incapable of pinning it — any bound at least
 as large passes. Such a constant needs its own numeric guard, and that guard's
 apparent triviality is a property of correct design, not a smell.
+## 2026-09-02 — E-THE-PERIPHERY-OF-A-STRATUM-IS-THE-OTHER-STRATA-1 — PROBE-HOUSE-DIFFERENTIAL-1 reported: the abstraction stratum carries the House cycle; the runner-up synthesis stratum fires on everything and moves nothing
+
+**Status:** FINDING (measured, `crates/lance-graph-planner/examples/house_differential.rs`, release, N = 200 arenas × 25 permutations) + CORRECTION of the plan's rung reading. **Confidence:** High on the measurement; the effect size is fixture-dependent (see robustness).
+
+**The correction first.** The probe's first cut sampled the periphery from `RungLevel::Counterfactual.peripheral_sample_where(k)` — empty, because no recipe has `min_rung` above `Counterfactual` — and reported condition (c) "structurally unrunnable". That was the scalar-era reading of `RungLevel` as one active position. The operator's ruling is the strata model (`persona-vs-rung-ladder.md`, 2026-08-30 scope correction): rungs are strata scheduled in parallel in dependency order. The periphery of the board stratum (RCR #4 / ASC #7, floor 6) is therefore **the other strata** — CAS #8 (floor 0) feeding it, CR #11 (floor 3) consuming it — not the set of recipes the board's own rung excludes. The plan's "ASC/CR open at `Analogical`" and its rung-elevation loop were the same misreading; both corrected in `house-differential-style-v1` §2/§3.
+
+**The measurement.** A0 (RCR alone) p@1 0.320; A1 / A1c (+admit, +ASC on the leader, +council) 0.610; A2 with the abstraction stratum S0 (`cas_abstract` on `C*`, CAS-down admits `{G→O_far, C*→G} ⊢ C*→O_far`, focus set = every observable `is_a` subject, never the label) 0.820; null (identical procedure, feature half shuffled, far parent re-owned, distinct `(cause, feature)` pairs enforced) mean 0.343, **p95 of the per-permutation aggregates 0.395**. (Numbers revised before merge after two review findings on #1141: the first cut focused `cas_abstract` on `C*` (Codex) — making it label-blind left the real arm unchanged and lowered the null from 0.475 to 0.425, the label focus had been starving the null; and the null shuffle admitted duplicate pairs that `observe` pools into stronger beliefs (CodeRabbit) — enforcing distinctness moved it to 0.395.) Planted cause eliminated 0/200. S0 fires on exactly the 100 far-fact arenas (0.500 — can-fire and can-stay-silent). **Variant 2: base path `S0 + board` PASS, council-gated S3 arm INCONCLUSIVE; variant 1 KILL** on (c): the synthesis stratum S3 (`cr_synthesize` of the runner-up's held-aside report) ran on 182/200 arenas, reordered the tail on 177, and moved p@1 on none (`A2-S3` ≡ `A1c`) — the "fires on everything" defect, in S3's design.
+
+**Why variant 2 is recorded as two results.** It gates S3 on the council's `split`; the council split on 0/200 arenas, so S3 never ran — the pre-registered rule returns PASS for the procedure, but what was exercised is S0 + board; the S3 arm is inconclusive until a fixture makes the split reachable. The cause is the probe's own signal derivation (`humility = 1 − margin(top1, top2)` ≈ 1, where `InnerCouncil::from_signals` gives Catalyst weight `1 − |humility − 0.5|·2 = 0`), not a council defect. Nothing about CR as a synthesis stratum is established either way.
+
+**Four things worth keeping.**
+1. The null's comparator is the p95 of the AGGREGATE statistic across permutations (25 values), not the p95 of 200 per-arena hit rates (0.720 here — a tail statistic of a different distribution). The first cut used the latter and would have compared 0.820 against the wrong number.
+2. A size-preserving shuffle must also shuffle what it does not touch: leaving the far parent owned by `C*` under the null leaks the planted identity (AN p95 0.475 → 0.735 on the pre-review cut). A2 still cleared the leaky null, by 0.085.
+3. A procedure that names the hidden label anywhere — even in a stratum that "only" abstracts — is not a procedure that can be applied; derive every focus set from what the arena shows (Codex P1 on #1141). An elimination read off a `(cause, expectation)` pair is "ranked low"; the predicate reads the belief's recorded contradiction (Codex P2), proven two-sided by G5. And a permutation null must preserve DISTINCTNESS, not just counts: a duplicate `(cause, feature)` pair is revised into a stronger belief by `observe`, which the real fixture never does (CodeRabbit) — 11 784 repair swaps over 5 000 shuffles says it was the common case.
+4. A stratum contributes when its chain carries evidence stronger than the direct band: far fact weakened to the rule-edge band (0.9/0.6) ⇒ S0 inert, A2 = A1c. The fixture strengths were pre-registered in the file before the official run; the dependence is a property of the mechanism, stated here so it is not read as a universal gain.
+
+**Disable-runs, verified red:** G1 (S0 forced on under `STRATA_OFF`), G4 (`CasUp` in place of `CasDown`), G5 (contradiction half dropped), jitter zeroed ⇒ A0 p@1 = 1.000 by tie-break.
+
+Refs: `house-differential-style-v1` §2 step 2, §3, §4 RESULT; `E-PERIPHERAL-DISSENT-GUARDS-THE-STRATIFICATION-1`; `E-A-GHOST-TRACE-IS-NOT-THE-COUNTERFACTUAL-LANE-1`; `nars/tactics.rs` `cas_abstract` (:359) / `cr_synthesize` (:503); `recipes.rs` `min_rung` (:505); D-HOUSE-1 → Shipped, D-HOUSE-2/3/5 unblocked on the stated terms, D-HOUSE-1d re-scoped to strata budgets.
+
 ## 2026-09-02 — E-THE-DTO-LADDER-IS-THE-ALU-BUS-AND-WAS-ALREADY-RULED-1 — a session re-asked a question `.claude/v3/` had answered on 2026-07-02, and proposed a probe on the wrong axis before reading it
 
 **Status:** CORRECTION + FENCE (same day, same branch, before merge). **Confidence:** High.

.claude/board/PR_ARC_INVENTORY.md

@@ -10,6 +10,16 @@
 > census §8.3 trap 10: read the body FIRST, then open for write — never
 > inline both in one expression.
 
+## 2026-09-02 — lance-graph branch `claude/medcare-rs-continue-6nhbxn` (D-HOUSE-1 PR, after #1139) — PROBE-HOUSE-DIFFERENTIAL-1 run and reported
+
+- **Added:** `crates/lance-graph-planner/examples/house_differential.rs` (the probe; no library surface): fixture, arms A0 / A1 / A1c / A2-S0 / A2-S3 / A2 (variant 1) / A2' (variant 2), size-preserving null with far-parent re-owning, guards G1–G4, per-variant verdict. Plan §4 RESULT; EPIPHANIES `E-THE-PERIPHERY-OF-A-STRATUM-IS-THE-OTHER-STRATA-1`.
+- **Changed:** plan §2 step 2 (periphery = the other strata), §2 closing paragraph (strata in dependency order, no elevation loop), §3 bullet 1 (per-stratum floors; ASC opens at `Counterfactual`, not `Analogical`), §4 arms / fixture / guards (as run).
+- **Measured:** A0 0.320 · A1c 0.610 · A2 0.820 · null p95 0.395 (aggregate; label-blind S0 and distinct-pair shuffle — 0.475 on the first cut that focused S0 on the label, 0.425 before duplicates were removed) · elim 0 (arena predicate; by construction for `C*`, G5 two-sided) · S0 fires 100/200 · S3 fires 182/200 with zero p@1 effect · council split 0/200. Variant 1 KILL on (c); variant 2: base path (S0 + board) PASS, council-gated S3 arm INCONCLUSIVE (never exercised). Guards G1/G4/G5 disable-verified red; fixture guard (jitter off ⇒ A0 = 1.000) verified. Robustness: weak far fact ⇒ S0 inert; leaky null ⇒ p95 0.735.
+- **Review folded in (pre-merge):** Codex P1 — S0 focus set derived from observable `is_a` subjects; Codex P2 — elimination predicate reads the belief's contradiction + floor; CodeRabbit — null shuffle enforces distinct `(cause, feature)` pairs (repaired shuffle, counts reported), variant 2 recorded as base-path PASS + S3 inconclusive, board/plan rows aligned.
+- **Locked:** the strata reading of the periphery for this plan; the aggregate-p95 comparator; far-parent re-owning in the null.
+- **Deferred:** D-HOUSE-2/3 unblocked by the base-path PASS, not started; D-HOUSE-5 half unblocked, still gated on the D-TSC-2/3 mint; D-HOUSE-1b/1c gated on this report as before; D-HOUSE-1d re-scoped to strata budgets before scheduling; a CR stratum with a reachable council gate is a new probe, not a re-run.
+- **Confidence:** High.
+
 ## 2026-09-02 — lance-graph branch `claude/medcare-rs-continue-6nhbxn` (D-TEH-1 PR, after #1138) — `bridge_gate` → contract; callcenter drops its thinking-engine dependency
 
 - **Added:** `crates/lance-graph-contract/src/bridge_gate.rs` (seven items + 9 tests, moved not re-derived); `pub mod bridge_gate` in the contract lib; `thinking_engine::bridge_gate` as a re-export shim keeping `pure_ops_dont_touch_gate` engine-side.

.claude/board/STATUS_BOARD.md

@@ -3,14 +3,14 @@
 | D-id | deliverable | status |
 |---|---|---|
 | D-HOUSE-0 | the plan: parts list of shipped surfaces (file-cited), the composition as a recipe-level style program, ladder placement, pre-registered falsifier | **Shipped (plan-only, this PR)** |
-| D-HOUSE-1 | PROBE-HOUSE-DIFFERENTIAL-1: planted-cause recovery on synthetic arenas with shared-predicate distractors; arms RCR-alone / +admit+ASC / +council control / full cycle / shuffle null (fixture direction `cause→feature` + `case→feature` so RCR yields `case→cause`; `quorum_project` NOT on the path — it is a `todo!`); PASS iff Δp@1 ≥ 0.05 above null p95, planted cause never eliminated, periphery fires on 10–90 % | **Queued — next step** |
+| D-HOUSE-1 | PROBE-HOUSE-DIFFERENTIAL-1: planted-cause recovery on synthetic arenas with shared-predicate distractors; arms RCR-alone / +admit+ASC / +council control / full cycle / shuffle null (fixture direction `cause→feature` + `case→feature` so RCR yields `case→cause`; `quorum_project` NOT on the path — it is a `todo!`); PASS iff Δp@1 ≥ 0.05 above null p95, planted cause never eliminated, periphery fires on 10–90 % | **Reported 2026-09-02** — variant 1 (S3 unconditional) KILL on (c), fire rate 0.910; variant 2: **base path (S0 + board) PASS** — A2 0.820 vs A0 0.320 vs null p95 0.395 (label-blind S0, distinct-pair shuffle), elim 0, fire 0.500 — and **council-gated S3 arm INCONCLUSIVE** (council split 0/200, S3 never ran; needs a fixture with a reachable split); periphery re-read as the other strata (plan §2/§3 corrected) |
 | D-HOUSE-1b | resonance prior probe (`res-seed` + `res-gate`): anchoring vs recognition, same arenas, ρ < 1 guard against the D-HOUSE-1 A2 ordering (plan §4b) | Gated on D-HOUSE-1 reported |
 | D-HOUSE-1c | signed-qualia contrast probe: one i4 per top-3 candidate from the ASC outcome into existing `QualiaI4_16D` lanes, post-challenge re-rank only, all-zero-signs must equal A2 (plan §4b) | Gated on D-HOUSE-1 reported |
-| D-HOUSE-1d | aperture hypothesis (the "humor" label): incongruity → `k` across passes; a hypothesis until its operator is defined and measured (plan §4b) | Not scheduled |
-| D-HOUSE-2 | discriminating-evidence readout (clone arena, observe hypothetical premise, report largest rank displacement) | Gated on D-HOUSE-1 PASS |
-| D-HOUSE-3 | elimination predicate over `truth`/`contradiction`/`stamp` + never-re-enter falsifier | Gated on D-HOUSE-1 PASS |
+| D-HOUSE-1d | aperture hypothesis (the "humor" label): incongruity → `k` across passes; a hypothesis until its operator is defined and measured (plan §4b) | Not scheduled — re-scoped 2026-09-02: the knob is the strata budgets, not `k` over an excluded set |
+| D-HOUSE-2 | discriminating-evidence readout (clone arena, observe hypothetical premise, report largest rank displacement) | Unblocked by the D-HOUSE-1 base-path PASS (2026-09-02); not started |
+| D-HOUSE-3 | elimination predicate over `truth`/`contradiction`/`stamp` + never-re-enter falsifier | Unblocked by the D-HOUSE-1 base-path PASS (2026-09-02); not started |
 | D-HOUSE-4 | anchoring alarm (free energy between wisdom-marker prior and evidence) | Gated on `thinking-engine-harvest-closure-v1` D-TEH-2 |
-| D-HOUSE-5 | the style→recipe edge (O2) with House as its first program | Gated on D-HOUSE-1 PASS and the D-TSC-2/3 mint |
+| D-HOUSE-5 | the style→recipe edge (O2) with House as its first program | D-HOUSE-1 half unblocked (2026-09-02, base-path PASS; the edge selects STRATA, CAS-down as the board's feeder); still gated on the D-TSC-2/3 mint |
 | D-HOUSE-6 | consumer pointer: `DiffRow` READ fields `eliminated` / `discriminator` (MedCare-rs) | pointer only, out of scope here |
 
 ## thinking-engine-harvest-closure-v1