D-CZ-1 PASSES, D-CZ-0 had no artifact behind it — plus the arc's rated formula matrix

.claude/board/EPIPHANIES.md

@@ -1,3 +1,100 @@
+## 2026-08-12 — E-A-FIGURE-CITED-TWICE-IS-NOT-CONFIRMED-ONCE-1
+
+**Status:** FINDING `[G]` — measured. Probe
+`probes/weather-p1/substrate_comfort_d_cz_0_1.py` + `.json`, one run,
+committed with the artifact this entry is about.
+
+**The defect.** `substrate-comfort-zones-v1` §1's nine-row regime preflight
+was marked **DONE** on `STATUS_BOARD.md`, quoted in the `PR_ARC_INVENTORY`
+entry, quoted again in the `LATEST_STATE` row, and its headline (a 9.3×
+dynamic range across four regimes) was carried into three PR bodies. **No
+script and no JSON producing any of those figures had ever been committed.**
+
+**The part that makes it worth an entry rather than a fix.** On #945 I ran
+an explicit self-audit *specifically because* this arc keeps shipping
+summary claims that do not survive checking — and reported the regime-ladder
+figures as verified. What that audit actually did was confirm the arc entry
+matched the PLAN. Both are prose. Neither is a measurement. **A figure cited
+in two documents has been cited twice, not confirmed once** — and an audit
+that compares two documents to each other measures their consistency, which
+is not the property anyone cares about.
+
+**The rule.** An audit must terminate at an ARTIFACT — a committed script, a
+committed JSON, a re-run — or it is a spell-check. Concretely: when a
+board row says DONE, the check is *"name the file that produced it"*, not
+*"does the summary match the plan"*.
+
+**What the reproduction then found**, none of which was visible from the
+prose:
+
+1. **Only 4 of 9 rows are reproducible at all.** R1/R2/R3 centres are in the
+   plan and R4's come from committed `comet_tail_f14.json` (the qualifying
+   filter — `status == "OK"` AND `displacement_km ≥ 250` — itself had to be
+   *recovered* by reproducing the stored `n_qualifying = 19`, and is now
+   asserted in code). The five EXCLUDED land candidates carry measured
+   figures whose **box centres were never written down anywhere**. They are
+   unreproducible as committed. No coordinates were invented to fake them.
+2. **The `|∇p|` definition was never committed either — only values.** Four
+   candidates were computed and the winner decided from the data: the
+   recorded figures are **Pa per grid cell with NO cos(lat) metric**
+   (max deviation 0.069; the next-best candidate is off by 0.398). It is a
+   plain `np.gradient` over the raw lat/lon array.
+3. **Consequence, bounded:** ignoring cos(lat) understates the ZONAL
+   gradient by `1/cos(lat)`, so R3 at 60 N is ~40 % low. Metric-corrected
+   the ladder reads 10.3 / 15.5 / 61.2 / 100.9 — the **ORDER survives** and
+   the dynamic range **widens** 9.3× → ≈9.8×. The regime axis stands; the
+   recorded magnitudes do not. Reported as a correction, not as a collapse.
+4. **A defect in the reproduction itself, caught by its own guard.** The
+   first pass measured all 19 storm boxes at the *preflight* timestep rather
+   than at each storm's own `t0` — "the places where storms once were, at an
+   unrelated hour". That inverted R3/R4 (220 vs 208) and would have read as
+   the regime ladder failing C1. Fetching per storm (19 extra chunks)
+   restores 37 / 56 / 220 / 363. Separately, a seam assertion fired on a
+   real storm centred at 353.4 E: a plain longitude slice there yields a
+   one-column box and a plausible, meaningless number. **Both were caught
+   because the probe asserted rather than assumed.**
+
+---
+
+## 2026-08-12 — E-THE-METRIC-THAT-SEPARATES-ONE-COMPARISON-IS-BLIND-TO-ANOTHER-1
+
+**Status:** FINDING `[G]` — measured, same run.
+
+**`ρ` is saturated on the diagonal and enormous off it.** Measured over four
+regimes at 256 palette levels on MSLP:
+
+| comparison | `ρ` separation |
+|---|---|
+| real arm vs real arm (`CAL-ABS` vs `CAL-RANK`, both own-calibration) | **3×10⁻⁶ … 4.7×10⁻⁵** |
+| real vs degraded (`CAL-SHUFFLE` 0.003–0.159, `GEO-DEGENERATE` 0.29–0.48) | **~0.5 … ~1.0** |
+
+Both real arms reconstruct the ordering essentially perfectly (ρ > 0.99996),
+because 256 levels on a smooth pressure field is a very fine quantization.
+So `ρ` has **four orders of magnitude of range for transfer loss `L`** (an
+off-diagonal quantity) and **none at all for C4** (which compares two real
+arms on their own diagonals). RMSE, which the same rewrite had just demoted,
+*does* separate the real arms: ratios 3.96 / 1.85 / 1.14 / 1.49.
+
+**The rule: a metric is not good or bad, it is good or bad FOR A GIVEN
+CONTRAST.** Choosing one for a plan as a whole — as the cross-swap rewrite
+did, elevating `ρ` and demoting RMSE in one move — is a category error the
+moment the plan contains two contrasts with different dynamic ranges. The
+resolution here is not to revert: `L` keeps `ρ`, C4 moves to RMSE in Pa with
+`ρ` as a floor check (< 0.999 = broken, not merely lost).
+
+**Why amending after data is legitimate here, and when it stops being.**
+D-CZ-1's *stated purpose* is to test the apparatus before the expensive
+cells, and **no C4 cell has been scored** — a metric found blind in
+preflight is exactly what preflight is for. It becomes illegitimate the
+instant one C4 measurement exists. Recorded as an amendment carrying its
+trigger (`§6.4`), never edited into the bar silently.
+
+**Cross-ref:** this is the same shape as #926's Fisher-z result (8.3× better
+in the storm tail on the raw field, 4.7× worse on ring means — one encoding,
+opposite verdicts by what it represents), which §0 already cites as a
+founding motivation for this plan. The plan's own primary metric turned out
+to be another instance of the phenomenon the plan was built to map.
+
 ## 2026-08-12 — E-A-HORSE-RACE-IS-NOT-A-CROSS-SWAP-1
 
 **Status:** FINDING `[H]` — methodological, operator-ruled. Caught on

.claude/board/STATUS_BOARD.md

@@ -1,3 +1,29 @@
+## SUBSTRATE_FORMULA_MATRIX — the arc's rated inventory (2026-08-12)
+
+Document: `probes/weather-p1/SUBSTRATE_FORMULA_MATRIX.md`. Not a plan and not
+a deliverable ladder — the **consolidated rating** of every formula,
+encoding, sampling geometry, physical model and statistical instrument the
+weather arc (#920–#946) actually put under a pre-registered bar.
+
+| D-id | Deliverable | Status | Feeds |
+|---|---|---|---|
+| D-MTX-1 | Re-extract every tested primitive from the COMMITTED artifacts (7 parallel readers + 1; 131 primitives, 2.35 M subagent tokens) | **DONE** | the matrix's provenance rule — built from artifacts, never from session memory |
+| D-MTX-2 | Two-scale rating (fitness A/B/C/D/V × evidence `[G]`/`[H]`/`[S]`), 56 rows across physics / encodings / geometries / instruments | **DONE** | reading comfort zones off instead of arguing them |
+| D-MTX-3 | Known-effect vs discovered-explanation pairing (14 pairs, K1–K14) | **DONE** | separates prior art from what this arc added — incl. 4 cases where measurement went AGAINST the prior |
+| D-MTX-4 | Figure verification against source JSONs (28 headline figures + circular/CT set) | **DONE — 0 mismatches**, 1 rounding fixed (Rayleigh p 0.689 → 0.688) | the audit-terminates-at-an-artifact rule, applied to this document itself |
+| D-MTX-5 | Refresh after the cross-swap matrix runs (C2–C6) | Queued | every off-diagonal cell is still unmeasured; §5 says so explicitly |
+
+**Why a "C" tier exists.** The arc's founding result is that most substrate
+formulas are neither good nor bad — they have homes. Fisher-z alone carries
+**three** measured verdicts (tail read 8.3× win, level read 4.7× loss,
+CI-frame not-a-win). A single ranking would have destroyed the finding.
+
+**What the matrix is honest about:** the majority of its rows are negative.
+Two entries are **V (VOID)** rather than D — the apparatus could not
+distinguish anything — and §5 lists 13 gaps including *every* off-diagonal
+cross-swap cell, all ten EV probes, and five preflight rows that are
+permanently unreproducible because their coordinates were never recorded.
+
 ## substrate-comfort-zones-v1 — the comfort-zone map (PRE-REGISTERED 2026-08-12)
 
 Plan: `.claude/plans/substrate-comfort-zones-v1.md`. A **cross-swap**
@@ -7,8 +33,8 @@ already run and it corrected two regime definitions before any bar existed.
 
 | D-id | Deliverable | Status | Feeds |
 |---|---|---|---|
-| D-CZ-0 | §1 regime preflight (`\|∇p\|` ladder, elevation-confound screen, speed-is-not-the-discriminator finding) | **DONE** — ladder R1 Amazon 10.2 → R2 ocean 14.9 → R3 W Siberia 43.8 → R4 storm 95.6 (9.3× range); 4 land candidates excluded on elev σ > 150 m | the regime axis all other rows score on |
-| D-CZ-1 | C0 controls (shuffled codebook + degenerate geometry), losability-smoke-tested BEFORE the full run | Queued | gates every cell — a control that can't lose voids its cell |
+| D-CZ-0 | §1 regime preflight (`\|∇p\|` ladder, elevation-confound screen, speed-is-not-the-discriminator finding) | **DONE — and now REPRODUCED + partly corrected (§6.1/6.2).** It had NO committed script or JSON when marked DONE. 4 rows reproduce (1.004/1.022/0.994/0.931); the 5 EXCLUDED land candidates are **unreproducible** (centres never recorded). Definition identified: Pa per grid cell **without cos(lat)** — R3 ~40 % low; metric-corrected ladder 10.3/15.5/61.2/100.9, order survives, range 9.3× → **9.8×**. Original text:**  — ladder R1 Amazon 10.2 → R2 ocean 14.9 → R3 W Siberia 43.8 → R4 storm 95.6 (9.3× range); 4 land candidates excluded on elev σ > 150 m | the regime axis all other rows score on |
+| D-CZ-1 | C0 controls (shuffled codebook + degenerate geometry), losability-smoke-tested BEFORE the full run | **DONE — PASS** (`substrate_comfort_d_cz_0_1.py/.json`). Both controls lose to both real arms on BOTH metrics in ALL FOUR regimes; mechanism visible (`GEO-DEGENERATE` saturates 92–97 %). C1b `separation` = **6.28** vs the ≥ 3 bar. **AND the run amended C4**: `ρ` is SATURATED on the diagonal (real-arm spread 3e-6…4.7e-5) so C4 could not have fired — `L` keeps `ρ` off-diagonal, C4 moves to RMSE in Pa (§6.4) | gates every cell — a control that can't lose voids its cell |
 | D-CZ-2 | C1 regime-ladder stability across ≥3 timesteps | Queued | anti-cherry-pick on the whole regime axis |
 | D-CZ-2b | **C1b constancy is measured** — `separation = (between-box range)/(mean within-box σ)` ≥ 3 | Queued | earns the phrase "held constant"; otherwise a caveat rides every cell |
 | D-CZ-2c | **C1c the suitability ASSUMPTION** — regimes must differ in autocorrelation decay + rank-distribution shape, not only in `\|∇p\|` | Queued | a null here VOIDS the cross-swap reading (spread manufactured on the wrong axis) |