D-TEH-3: calibration math → jc (drift, quorum, lift-gate comparisons); lab cronbach.rs deleted

.claude/board/EPIPHANIES.md

@@ -1,3 +1,21 @@
+## 2026-09-02 — E-THE-LIFT-GATE-FOUND-A-TIE-BLIND-SPEARMAN-1 — the D-TEH-3 comparison on a distinguishing fixture separated a same-formula copy from a wrong-estimator copy, and the tie-free fixtures the lab had used could not have
+
+**Status:** FINDING (measured, tests committed in `crates/jc/src/reliability.rs`). **Confidence:** High.
+
+**What the ruling asked.** `E-JC-IS-THE-HOME-OF-ALL-CALIBRATED-MATH-1`: a calibration routine outside jc is LIFTED if correct, or the jc version is PERFECTED and the copy dies — decided on a fixture that can distinguish the two implementations. D-TEH-3 ran that gate on the thinking-engine battery. Four verdicts, two of them not the one a code read would have given.
+
+**Cronbach α — LIFT, with a precision caveat the fixture had to expose.** `thinking_engine::cronbach::cronbach_alpha` and `jc::reliability::cronbach_alpha` are the same estimator (population variances, `k/(k−1)·(1 − Σσ²/σ²_total)`); on the known-value fixture they agree to `1e-5` (0.984615). That agreement proves nothing by itself, so the same fixture was shifted by `1e7`: α is affine-invariant in exact arithmetic and the `f64` estimate does not move (`< 1e-9`), while the retired `f32` form loses the fixture outright (its mean rounds to the nearest unit at that magnitude and the deviations carry no information). The copy dies; jc's `f64`/`Option` form is the survivor, not because it is "newer" but because the fixture showed which one survives the data.
+
+**Spearman ρ — PERFECT-IN-JC, and the lab's own tests were structurally blind to it.** `ground_truth::spearman_rank_correlation` ranked ties by sort POSITION (index order), not by average rank. On tie-free data it agrees with jc to `1e-6` — and every one of its three retired unit tests was tie-free. On `x=[1,2,3,4]`, `y=[1,2,2,3]` it returns ρ = 1.000 (a "perfect monotone match"); the average-rank form returns the textbook 0.948683. A calibration battery that compares baked `u8` lens distances against ground-truth cosines will hit ties constantly (256 levels over thousands of pairs), so the retired form was systematically optimistic exactly where it was used. jc already carried the tie correction; nothing was added to it except the comparison test that keeps the retired form on record.
+
+**Re-encode drift and the correction-delta summary — LIFTED as-is into `jc::drift`.** The convergence statistic (`|e_i − e_{i−1}| < 1e-15` against the ORIGINAL value, tail-filled history, `safe = converged_at < max_iterations`) and the delta battery (mean / mean|δ| / max|δ| / population σ / two threshold fractions) were correct; what changed is their shape — the codec choice (BF16, γ+φ, the full chain) is now a round-trip closure the lab passes in, and the two cut-offs (0.01 / 0.1) are parameters the lab names (`MATERIAL_CORRECTION` / `LARGE_CORRECTION`) instead of constants buried in the statistic. The x256 re-encode proof (14 lab tests) runs unchanged through jc. **The lens quorum** (per-pair `1 − σ/σ_max` over `u8` tables, the 230/179/128 bands, the per-subject Cronbach report) went to `jc::quorum` the same way; it never was an α per pair and its docs now say so in the type name.
+
+**The generalizable point.** "Agrees on the existing tests" was true for both copies and meant nothing: the tests were the fixtures on which the implementations coincide. The gate had to be run on the input class each estimator would actually see (large offsets for `f32`, ties for ranks). A lift decision made from a code read would have kept the tie-blind Spearman — it looks like the textbook formula.
+
+**Still outside jc, recorded rather than swept.** `ndarray::hpc::reliability` (tie-aware, `f64`, returns `0.0` on degeneracy) and `perturbation-sim::stats` (a deliberate zero-dep mirror of it) carry the same four estimators. Both are correct in the sense above; neither is jc. Filed as `TD-RELIABILITY-COPIES-OUTSIDE-JC-1` — the ruling's own text names the ndarray direction ("once ndarray is proven bit-exact the same math is re-importable from jc"), which is a deliberate PR with a bit-exactness gate, not a drive-by.
+
+Refs: `E-JC-IS-THE-HOME-OF-ALL-CALIBRATED-MATH-1`; `thinking-engine-harvest-closure-v1` §1d / §3 W2 / §5 D-TEH-3; `crates/jc/src/{reliability,drift,quorum}.rs`; the falsifiability rule (CLAUDE.md § The falsifiability rule).
+
 ## 2026-09-02 — E-THE-CALIBRATION-GATE-REVERSED-THE-DECLARED-FLOOR-1 — D-TEH-2: the ghost prior landed in the planner, and the pre-registered gate overruled the floor its author had declared
 
 **Status:** FINDING (measured, `crates/lance-graph-planner/src/nars/ghost_prior.rs`, test `calibration_gate_picks_the_default_floor`). **Confidence:** High on the fixture; the fixture is one recurrence shape (256 atoms, a 3-atom pattern, 0–30 stale patterns, ages 0–60).

.claude/board/LATEST_STATE.md

@@ -1,3 +1,13 @@
+## 2026-09-02 — branch (D-TEH-3, after #1142 merged): calibration math lifted into jc — INVENTORY DELTA
+
+- MERGED #1142 (D-TEH-2, `3c5f040`): no contract type; planner `nars/ghost_prior.rs` + lab-crate edits + `weather-poc.yml` sibling checkout (see PR_ARC).
+- ADDED `jc::drift` — `ReencodeDrift` / `reencode_drift(value, max_iterations, codec, round_trip: FnMut(f32) -> f32)` / `DriftBatch` / `reencode_batch` (the re-encode convergence statistic, codec-agnostic), `DeltaSummary` / `delta_summary(deltas, material_threshold, large_threshold) -> Option` / `DeltaSummary::empty`, `CONVERGENCE_EPS`. 8 tests (identity converges at 1 with a filled tail; multiplicative drift never converges with a strictly growing history; damped codec converges late and the budget binds two-sided; zero budget proves nothing; batch worst/count; hand-computed summary; thresholds load-bearing both ways; empty / non-finite → `None`).
+- ADDED `jc::quorum` — `pairwise_agreement_u8(tables, n) -> Option<Vec<u8>>` (per-pair `1 − σ/σ_max(k)`, `max_u8_variance(k)` = `255²·⌊k/2⌋⌈k/2⌉/k²` — the lifted source used the even-`k` ceiling for every `k`, so a maximally split 3-lens cell scored 15 instead of 0; perfected in jc per the ruling, Codex on the lift PR), `QuorumLevel { High, Medium, Low, Ambiguous }` with `HIGH_FLOOR`/`MEDIUM_FLOOR`/`LOW_FLOOR` = 230/179/128 and `needs_leaf_validation`, `CronbachReport` / `cronbach_report(items) -> Option` (α by delegation to `reliability::cronbach_alpha` + per-subject variances + the `mean + σ` disagreement count). 7 tests (coincidence → 255 off-diagonal too; odd-`k` `[0,0,255]` → 0 and `[0,0,128]` strictly between; 0-vs-255 → 0 and a 100-vs-150 pair → 205; degenerate → `None`; both sides of every band floor; the one disputed subject is the one flagged, with a mild subject above the mean but under `mean + σ` as the anti-vacuity for the disable; `None` where α is).
+- ADDED `jc::reliability` tests only — the two D-TEH-3 lift-gate comparisons carrying the retired lab forms verbatim (`retired_cronbach_f32`, `retired_spearman_ordinal`): cronbach agrees on the known-value fixture and loses the `1e7`-shifted one; spearman agrees tie-free and diverges on one tie (1.000 vs 0.948683). No estimator changed. jc lib: 135/135.
+- REMOVED (lab crate, excluded) `thinking_engine::cronbach` (`cronbach_alpha`, `variance_agreement_scores`, `QuorumLevel`, `CronbachResult`, `cronbach_analysis`) and `ground_truth::{spearman_rank_correlation, ranks}` with their tests. `reencode_safety.rs` is now glue over `jc::drift` (`pub type ReencodeSafety = jc::drift::ReencodeDrift`; the three codec wrappers pass round-trip closures; `test_reencode_batch` keeps its tuple shape; `test_zipper_offsets` + all 14 tests unchanged and green). `silu_correction.rs`: `pub use jc::drift::DeltaSummary as CorrectionStats`, `correction_stats -> Option<CorrectionStats>` is an adapter with the cut-offs named `MATERIAL_CORRECTION = 0.01f32` / `LARGE_CORRECTION = 0.1f32` (kept in the samples' precision and promoted with them, so a correction of exactly `0.1f32` stays "not more than" 0.1 — Codex); `None` = non-finite data (an invalid run), `Some(empty)` = genuinely empty (Codex: the first cut folded both into empty). `ground_truth::calibration::spearman_vs_ground_truth` calls `jc::reliability::spearman` (keeps its `0.0` fallback). `examples/certify_jina_v5_7lane.rs` calls `jc::reliability::cronbach_alpha` (`None` → `NaN`, which fails every `>=` verdict honestly). Lab `Cargo.toml` gains `jc = { path = "../jc" }`; lab lock adds `jc` and drops entries cargo no longer needs. Lab: `cargo check --lib --examples` clean except the pre-existing `tts_stream_hhtld` example break (bgz-tensor private fn, untouched, `TD-THINKING-ENGINE-EXCLUDED-DEBT-1`); lab tests for the three modules 23/23.
+- NOT IN THIS DELTA: the `semantic_chunker` / `spiral_segment` halves of D-TEH-3 (falsifier-gated, separate); `silu` and `cosine_f32` stay in the lab (activation + vector glue, not calibrated math); `ndarray::hpc::reliability` and `perturbation-sim::stats` copies (`TD-RELIABILITY-COPIES-OUTSIDE-JC-1`).
+- UNCHANGED: every contract type; every existing jc estimator's arithmetic, signature and semantics; the ALU artery files.
+
 ## 2026-09-02 — branch (D-TEH-2, after #1141 merged): ghost prior harvested into the planner — INVENTORY DELTA
 
 - MERGED #1141 (D-HOUSE-1, `3102776`): no contract type; planner example `house_differential.rs` + plan/board records (see PR_ARC).

.claude/board/PR_ARC_INVENTORY.md

@@ -10,6 +10,17 @@
 > 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-TEH-3 PR, after #1142) — calibration MATH → jc; lab copies deleted
+
+- **Added:** `crates/jc/src/drift.rs` (`ReencodeDrift`, `reencode_drift`, `DriftBatch`, `reencode_batch`, `DeltaSummary`, `delta_summary`, `CONVERGENCE_EPS`; 8 tests), `crates/jc/src/quorum.rs` (`pairwise_agreement_u8`, `max_u8_variance`, `QuorumLevel` + floors, `CronbachReport`, `cronbach_report`; 7 tests), two lift-gate tests in `crates/jc/src/reliability.rs`; `pub mod drift; pub mod quorum;` in jc `lib.rs`. EPIPHANIES `E-THE-LIFT-GATE-FOUND-A-TIE-BLIND-SPEARMAN-1`; TECH_DEBT `TD-RELIABILITY-COPIES-OUTSIDE-JC-1`.
+- **Removed:** `crates/thinking-engine/src/cronbach.rs`; `spearman_rank_correlation` + `ranks` (+3 tests) from `ground_truth.rs`; the private statistic bodies in `reencode_safety.rs` and `silu_correction.rs`.
+- **Changed (lab crate):** `Cargo.toml` (+`jc` path dep) and `Cargo.lock`; `lib.rs` (no `cronbach`); `reencode_safety.rs` (glue over `jc::drift`, tests verbatim); `silu_correction.rs` (`CorrectionStats` = `jc::drift::DeltaSummary`, named cut-offs); `ground_truth.rs` (calibration glue calls jc); `examples/certify_jina_v5_7lane.rs` (jc α, `Option` → `NaN`). Plan §5 D-TEH-3 → math half Shipped, §3 W2 result addendum; STATUS_BOARD D-TEH-3.
+- **Measured:** cronbach — same formula, agree to `1e-5` on the 0.984615 fixture, `f32` copy loses the `1e7`-shifted fixture while `f64` moves `< 1e-9` → LIFT; spearman — retired form ranked ties by position, agrees `1e-6` tie-free, returns 1.000 vs jc 0.948683 on `[1,2,2,3]` → PERFECT-IN-JC (already there); drift + delta summary + quorum → lifted as-is with the codec / cut-offs turned into parameters. jc lib 135/135, new modules clippy `-D warnings` clean, fmt clean; lab 23/23 on the three modules, `check --lib --examples` clean but for the pre-existing `tts_stream_hhtld` break.
+- **Review folded in (pre-merge, Codex P2 ×3):** odd-`k` quorum normalised by its own attainable variance (`max_u8_variance(k)`; `[0,0,255]` at `k = 3` scored 15 under the lifted even-`k` ceiling, now 0 — a defect the lift PERFECTED rather than carried); the lab cut-offs kept as `f32` and promoted with the samples (exactly `0.1f32` is not `> 0.1`); `correction_stats` returns `Option` — `None` for non-finite corrections, `Some(empty)` only for an empty set.
+- **Locked:** jc is the only home of the four estimators inside lance-graph; the lab carries codec/cut-off choices, never a statistic; the retired forms stay in jc's tests as the comparison record.
+- **Deferred:** D-TEH-3's `semantic_chunker` / `spiral_segment` halves (falsifier-gated); the ndarray / perturbation-sim copies (TD, deliberate PR with a bit-exactness gate); the pre-existing lab example break (D-TEH-5 / TD-THINKING-ENGINE-EXCLUDED-DEBT-1).
+- **Confidence:** High.
+
 ## 2026-09-02 — lance-graph branch `claude/medcare-rs-continue-6nhbxn` (D-TEH-2 PR, after #1141) — ghost prior harvested into the planner; `ghosts.rs` deleted
 
 - **Added:** `crates/lance-graph-planner/src/nars/ghost_prior.rs` — `GhostPrior`, `PriorFloor`, `Trace`, `calibration::{recurrence_fixture, discrimination}` (both `Option` below `FIXTURE_MIN_ATOMS` and above `FIXTURE_MAX_STALE_PATTERNS`), 15 tests (monotone decay to each floor with the inert cycle derived from the constant, decay-constant load-bearing both ways, two-sided free energy under both floors, the calibration gate on the declared default, anti-vacuity that the floors differ, prediction shape, imprint cap, echo rule, independence of two priors); `pub mod ghost_prior` + re-export in `nars/mod.rs`. EPIPHANIES `E-THE-CALIBRATION-GATE-REVERSED-THE-DECLARED-FLOOR-1`.

.claude/board/STATUS_BOARD.md

@@ -20,7 +20,7 @@
 | D-TEH-0 | census: live footprint measured (one required consumer via `bridge_gate`; one optional via `with-engine`), 51-file fate table, open-row reconciliation, four-wave closure, idea harvest, four rulings asked | **Shipped (plan-only, this PR)** |
 | D-TEH-1 | W1: `bridge_gate` (seven items) → `lance_graph_contract::bridge_gate`; callcenter re-imports and drops the path dep; thinking-engine keeps a re-export shim | **Shipped 2026-09-02** (before: required dep, 6 crossing sites, dep-drop alone fails 6 × E0433; after: zero thinking-engine deps in callcenter metadata, contract 1303/1303, callcenter 156/156, driver default + `with-engine` green, ALU artery files byte-identical). `with-engine` re-point deferred by stop condition: D-TTV-1 not landed |
 | D-TEH-2 | W2: ghost prior → planner `nars/ghost_prior.rs` over `WisdomMarker`, per-thought, two-sided falsifiers; crate `ghosts.rs` deleted | **Shipped 2026-09-02** — `GhostPrior`/`PriorFloor` in planner `nars/ghost_prior.rs`, 14 tests incl. the calibration gate (default floor = `Marker`, the gate reversed the first declaration); `ghosts.rs` + `think.rs` deleted; TD-GHOST-ECHO-DUP-1 resolved; consumer D-HOUSE-4 unblocked |
-| D-TEH-3 | W2: calibration MATH → `jc` (ruling 4, 2026-09-02: lift if correct, perfect in jc if not; crate copies die); glue stays in the lab crate; `semantic_chunker` / `spiral_segment` decided by their falsifiers | Queued |
+| D-TEH-3 | W2: calibration MATH → `jc` (ruling 4, 2026-09-02: lift if correct, perfect in jc if not; crate copies die); glue stays in the lab crate; `semantic_chunker` / `spiral_segment` decided by their falsifiers | **Math half Shipped 2026-09-02** — `jc::drift` (re-encode drift + delta summary) and `jc::quorum` (pairwise agreement, bands, Cronbach report) added; cronbach LIFTED (same formula; `f32` copy loses the `1e7`-shifted fixture), spearman PERFECTED-IN-JC (retired copy was tie-blind: 1.000 vs 0.9487 on one tie); lab `cronbach.rs` deleted, `reencode_safety` / `silu_correction` / `ground_truth` are glue over jc. Open: the `semantic_chunker` / `spiral_segment` falsifier halves |
 | D-TEH-4 | W3: ENTROPY M8 engine collapse with dtype parity suite; 5 cascade shapes + 3 lens modules collapse | Queued |
 | D-TEH-5 | W4: residue deleted, crate renamed `thinking-lab` with a `--manifest-path` CI line; §2 rows closed; `TD-THINKING-ENGINE-EXCLUDED-DEBT-1` paid | Queued — closes the chapter |
 

.claude/board/TECH_DEBT.md

@@ -1,3 +1,29 @@
+## TD-RELIABILITY-COPIES-OUTSIDE-JC-1 (2026-09-02) — OPEN
+
+**Two further copies of the Pearson / Spearman / Cronbach α / ICC battery live
+outside `jc`.** Found while running the D-TEH-3 lift gate
+(`E-THE-LIFT-GATE-FOUND-A-TIE-BLIND-SPEARMAN-1`):
+
+- `ndarray::hpc::reliability` (`pearson`, `spearman`, `cronbach_alpha`,
+  `icc_a1`, `FidelityReport`) — tie-aware, `f64`, returns `0.0` on degenerate
+  input where jc returns `None`. The PRODUCTION copy; consumed by the
+  edge-codec fidelity work.
+- `perturbation-sim::stats` — a self-described zero-dep mirror of the ndarray
+  copy so that crate's validation harness stays standalone.
+
+Neither is wrong the way the retired thinking-engine Spearman was (both average
+ranks over ties). They are still two more sources of truth for math the ruling
+says has ONE home. The ruling itself names the direction for the first — *once
+ndarray is proven bit-exact the same math is re-importable from jc* — so the
+pay-down is a deliberate PR with a bit-exactness gate (same fixtures as jc's
+lift-gate tests, plus the `0.0`-vs-`None` degeneracy contract decided
+explicitly), never a drive-by delete. The perturbation-sim mirror follows
+whatever the ndarray decision is.
+
+Owner: whoever next touches `ndarray::hpc::reliability` or `perturbation-sim`.
+Refs: `E-JC-IS-THE-HOME-OF-ALL-CALIBRATED-MATH-1`, `crates/jc/src/reliability.rs`
+(lift-gate tests), `thinking-engine-harvest-closure-v1` §1d.
+
 ## TD-PILLAR11-SCIENTIFIC-LOOPS-BYPASS-NDARRAY-SIMD-1 (2026-09-02) — OPEN
 
 **The debt is not missing SIMD support. It is scientific code bypassing the