AdaWorldAPI · AdaWorldAPI · Aug 12, 2026 · Aug 12, 2026 · Aug 12, 2026 · Aug 12, 2026
diff --git a/.claude/board/EPIPHANIES.md b/.claude/board/EPIPHANIES.md
@@ -1,3 +1,92 @@
+## 2026-08-12 — E-THE-DISPLACEMENT-FILTER-ATE-THE-STRANDED-STRATUM-1
+
+**Status:** FINDING `[G]` — W6 RUN (`comet_tail_w6.py`/`.json`), audited
+(raw predictors committed per storm after a self-caught gap), both a
+sample-composition finding and a genuine model-level negative result.
+
+**The mechanism-level result, stated first because it is the load-bearing
+one.** Report §10.2's dipole vector-sum model (`D = c_geo·P_geo +
+c_bow·P_bow`, a background-high neighbor predictor plus a relative-motion
+bow-wave predictor) was fit on CT-F14's 19 stored storms and **VOIDED by its
+own pre-registered anti-vacuity control**: single-geo R²=−0.104 (worse than
+predicting the mean); the permuted-P_bow control R²=−0.071 and the
+rotated-90° control R²=−0.062 BOTH exceed `single-geo + 0.03 = −0.074` — two
+deliberately wrong references score as well as or better than the real
+predictor. This is not a marginal miss — it is the anti-vacuity control
+doing exactly its job: rejecting a fit that has nothing to identify.
+
+**Checked before concluding it was a clean negative, per the standing
+measurement-skeptic discipline: no implementation bug found in the FIT
+itself, but a sign convention AND a units error were found in the
+NARRATIVE around it (codex + CodeRabbit P2/Major on PR #940, both real,
+fixed before merge).** One storm was independently re-fetched and hand-
+audited; extended to the full sample by committing the raw
+`D`/`P_geo`/`P_bow`/`A_H`/`d_H`/`v_rel_ms` per storm. Every value is
+physically sane (`A_H > 0` always, `d_H` inside the 600–2500 km annulus
+always, `v_rel` 2.8–27.9 m/s).
+
+**Sign:** `spine()`'s raw fit coefficient points toward the storm's HIGH
+side (the gradient of increasing residual pressure), while `P_geo`/`P_bow`
+both point toward the LOW side by construction — the exact convention
+`low_pole_bearing()` makes explicit via its own `(ph + π) % (2π)` flip. `D`
+is correctly `−spine(...)`; the first draft used the unflipped `coef`.
+Corrected: **`c_geo = +0.407` (the physically predicted positive sign —
+CORRECT)**, `c_bow = −0.0006` km⁻¹ (predicted positive — **wrong sign, but
+small**). Verified algebraically and numerically that this flip changes
+NOTHING about R² or the B0/B1 VOID verdicts (OLS is odd-symmetric in the
+fit target) — only the coefficient signs and the sentence describing them.
+
+**Units:** `D`/`P_geo` are [Pa/km]; `P_bow` is [Pa] — `c_geo` is
+dimensionless, `c_bow` carries km⁻¹, and OLS coefficients rescale inversely
+under column rescaling while R²/fitted-values stay fixed. **Raw `|c_bow|`
+was never valid evidence of "no measurable weight"**, and comparing
+`|P_bow|` to `|D|` directly (147×, as first reported) compounded the same
+mistake — Pa is not comparable to Pa/km at all. The dimensionally valid
+measure is the fitted CONTRIBUTION `|c_bow·P_bow|` against `|D|`, both in
+Pa/km: mean `|D|`=0.745, mean `|c_geo·P_geo|`=0.186 (25 % of `|D|`), mean
+`|c_bow·P_bow|`=0.068 (9 % of `|D|`) — the geo contribution is ~2.7× the
+bow contribution, MODEST rather than "no weight," and both remain
+consistent with the R²<0 finding that neither predictor meaningfully
+explains `D`'s variance.
+
+**The finding that generalizes past this one probe: an anti-vacuity control
+can be voided by SAMPLE COMPOSITION, and the reason is arithmetic, not
+physics.** B3's stranded-vs-moving stratification (`|v_storm| < 8 m/s`,
+the report's own named test of the "stranded-rescue" reading) came back
+**n=0 for the stranded stratum** — every one of the 19 storms has
+`|v_storm| ≥ 12.54 m/s`. This is not a null result about storm motion; it is
+a DIRECT ARITHMETIC CONSEQUENCE of CT-F14's own qualifying filter
+(`displacement_km ≥ 250` over the 6 h window): `250 km / 6 h = 11.574 m/s`,
+a hard floor on `|v_storm|` for ANY storm admitted to the sample. **A filter
+built to select clearly-moving storms for a displacement-scoring test
+silently and permanently excludes the storms a LATER, differently-motivated
+test (stranded-rescue) needs to see.** The stranded-rescue claim is
+therefore **UNTESTABLE on this sample, not refuted** — the untestability was
+knowable from the filter's own arithmetic before a single storm was fetched,
+and wasn't checked until B3 came back empty.
+
+**Consequence, stated as a reusable rule:** before scoring ANY new
+hypothesis against an EXISTING filtered sample, check whether the sample's
+own selection criterion is compatible with the new hypothesis's own
+discriminating variable — arithmetically, not by running the probe and
+discovering an empty stratum after the fact. A filter selected for one
+purpose (fast, clearly-displaced storms, easy to center-find and score
+against displacement) is not neutral with respect to every future question;
+it is a specific cut through the underlying population, and every later
+probe inherits that cut whether or not it is the cut that probe needs.
+
+**Consequence for the report and for CT-F17.** The report's §10.2 vector-sum
+model, AS SPECIFIED, is disconfirmed on this sample — not "unproven," not
+"needs more data" in the ordinary sense, but VOID by its own control. CT-F17
+(the fresh-sample verdict, gated on W6's result + an independent adversarial
+audit) is now moot **for this form of the model** — a fresh-sample test of a
+model that already fails identifiability on the stored sample is not the
+next useful step. Any REVISED form of the vector-sum model (multiple
+neighbors, a nonlinear bow term, per-storm coefficients) would need its own
+W6-shaped mechanistic test before earning a CT-F17 slot; a genuine
+stranded-rescue test needs a sample built without (or explicitly retaining
+slow storms despite) a displacement floor.
+
 ## 2026-08-12 — E-ON-A-GOLDEN-LATTICE-LOCALITY-IS-FIBONACCI-MEMBERSHIP-1
 
 **Status:** FINDING `[G]` — W5 RUN (`spiral_adi_probe.py`/`.json`) + a

diff --git a/.claude/board/PR_ARC_INVENTORY.md b/.claude/board/PR_ARC_INVENTORY.md
@@ -1,3 +1,71 @@
+## 2026-08-12 — the W6 dipole-deconvolution RUN lands — VOID by its own anti-vacuity control, plus a reusable sample-composition lesson (PR pending)
+
+- **Added.** `comet_tail_w6.py`/`.json` — the report §10.2 vector-sum
+  model (`D = c_geo·P_geo + c_bow·P_bow`) fit on CT-F14/F16's 19 stored
+  storms, per `weather-w-probes-v1.md` §3's pre-registered bars. New
+  epiphany `E-THE-DISPLACEMENT-FILTER-ATE-THE-STRANDED-STRATUM-1`.
+- **Locked — B0 VOIDS, and the model has nothing to identify.**
+  Single-geo R²=−0.104 (worse than the mean); both anti-vacuity controls
+  (permuted P_bow, P_bow rotated +90°) score at −0.071/−0.062, both
+  clearing the ≤`single-geo + 0.03` bar's ceiling of −0.074. B1/B2
+  correctly report VOID (B0 failed), not their own numbers, per the
+  pre-registered rule.
+- **Locked — checked for an implementation bug before calling it a clean
+  negative, per the standing measurement-skeptic discipline; found none
+  in the fit — but found a sign convention error AND a units error in
+  the FIRST DRAFT's narrative (codex + CodeRabbit P2/Major on PR #940,
+  both fixed before merge).** One storm independently re-fetched and
+  hand-audited (t0=54358); extended to all 19 by committing the raw
+  predictors per storm (the first run shipped only derived bearings — a
+  self-caught audit gap, fixed and re-run before any reviewer needed to
+  find it). **Sign:** `spine()`'s raw coefficient points toward the
+  storm's HIGH side (increasing residual), while `P_geo`/`P_bow` both
+  point toward the LOW side by construction — `D = −spine(...)` matches
+  the same `(ph+π)` convention `low_pole_bearing()` already uses.
+  Corrected: `c_geo = +0.407` (physically predicted positive — CORRECT,
+  the first draft's "−0.41, wrong sign" had the polarity backward);
+  `c_bow = −0.0006` km⁻¹ (predicted positive — wrong sign, small).
+  Verified algebraically and numerically that the flip changes NOTHING
+  about R²/B0/B1 (OLS is odd-symmetric in the fit target) — only the
+  coefficient signs and their description. **Units:** `D`/`P_geo` are
+  Pa/km, `P_bow` is Pa, so `c_geo` is dimensionless but `c_bow` carries
+  km⁻¹ — raw `|c_bow|` (or the first draft's `|P_bow|`/`|D|` ≈ 147×
+  comparison) was never valid evidence of "no measurable weight," since
+  OLS coefficients rescale inversely under column rescaling. The
+  dimensionally valid measure is the fitted contribution
+  `|c_bow·P_bow|` vs `|D|`, both Pa/km: mean `|D|`=0.745, geo
+  contribution 25 % of `|D|`, bow contribution 9 % of `|D|` — modest,
+  not "no weight," and consistent throughout with R²<0.
+- **Locked — the reusable lesson: B3's stranded stratum is EMPTY (n=0),
+  and the reason is arithmetic, not physics.** `min(|v_storm|) = 12.54
+  m/s` across all 19 storms — CT-F14's own qualifying filter
+  (`displacement_km ≥ 250` over 6 h) mathematically implies `|v_storm| ≥
+  250 km / 6 h = 11.574 m/s` for ANY admitted storm. **A filter selected
+  for one purpose (fast, cleanly-displaced storms) is not neutral for a
+  LATER, differently-motivated test (stranded-rescue) — the
+  incompatibility is knowable from the filter's own arithmetic before a
+  single fetch, and wasn't checked until the stratum came back empty.**
+  Filed as a reusable pre-flight check for any future probe reusing an
+  existing filtered sample.
+- **Locked — consequence for CT-F17.** The vector-sum model AS
+  SPECIFIED is disconfirmed on this sample, void by its own control —
+  not "unproven," not "needs more data." CT-F17's gate (W6's result +
+  independent adversarial audit) is now moot **for this form of the
+  model**: a fresh-sample verdict on a model that fails identifiability
+  on the stored sample is not the useful next step. A REVISED model
+  form needs its own W6-shaped mechanistic test first; a genuine
+  stranded-rescue test needs a sample built without the displacement
+  floor.
+- **Docs.** `weather-w-probes-v1.md` §3 gets the full RUN section.
+  `STATUS_BOARD.md` D-W6 moves from Queued to the RUN verdict.
+- **Confidence.** High — every figure re-derived from the committed
+  JSON (which now carries the raw per-storm predictors, not only the
+  fit summary) or from a live re-fetch audit, not asserted from the
+  first run's output alone.
+
+**Status:** RUN complete; awaiting a PR to land these results.
+Probe + docs — zero product code.
+
 ## 2026-08-12 — lance-graph #932 (MERGED) — the golden-ratio index floor + the temperament mechanism (merged before its own review findings could be addressed — see #933)
 
 - **Added.** The operator-ruled golden-ratio index floor (convergent index

diff --git a/.claude/board/STATUS_BOARD.md b/.claude/board/STATUS_BOARD.md
@@ -27,7 +27,7 @@ Wave 1 = parallel, no operator gate beyond go-ahead; gated rows named.
 |---|---|---|---|---|
 | D-W5 | Spiral-ADI anisotropy (v2, full-band control, V-matched iterations, bump 3.35σ from mask) | 1 | **B2 FAIL / B3 VOID CONFIRMED / B4 INCOMPLETE** — B2: real diffusion resolved, aniso 1.5251 vs 1.25 bar, clean baseline 1.0046, operator contributes ~0.52. B3: family A 99.68 % + family B 99.56 % (both link families, QUALIFYING population n=4.78M out of the headline lattice N=7.65M, not the 62k sub-sample) land on a pure Fibonacci offset — dominated by the two discovered strides 2584=F(18)/4181=F(19) respectively. B4: downgraded to a DESCRIPTIVE reading over n=8–17 only — n=19 was dropped without pre-authorization, bar not satisfied, stays open | domino.rs gather-design claim REFUTED at this test point (v1's "unblocked" was the same inert-operator artifact that also drove B2's false PASS); `E-ON-A-GOLDEN-LATTICE-LOCALITY-IS-FIBONACCI-MEMBERSHIP-1` strengthened, third dated update; B4's n=19 remains an open follow-up |
 | D-W2sA | Golden-vs-grid pairing on real cos-lat metric (zero-ties G1, CV G2) | 1 | **RUN — G1 VOID, G2/G4 FAIL (control degenerate: two identical translated grids are symmetry-uniform, CV ~1e-12 — cannot lose any evenness comparison; diagnosed via smoke test, run as-specified, `E-A-CONTROL-THAT-CANNOT-LOSE-IS-NO-CONTROL-1`)** | honest falsifier for evenness DEFERRED (offset/rotation-varied or spacing-mismatched control); §10.5 properties 1–3 untouched |
-| D-W6 | Two-component deconvolution (geo + bow, global lstsq, 38 eqs / 2 params; B3 = stranded stratification via v_rel) | 1 | Queued | dipole vector-sum identification; F17 gate |
+| D-W6 | Two-component deconvolution (geo + bow, global lstsq, 38 eqs / 2 params; B3 = stranded stratification via v_rel) | 1 | **RUN COMPLETE — B0 VOID** (single-geo R²=−0.104, worse than the mean; both anti-vacuity controls score ≥ the ceiling — the model has nothing to identify); B1/B2 correctly report VOID per rule; B3 stranded stratum EMPTY (n=0 — a structural consequence of CT-F14's displacement≥250km filter implying `\|v_storm\|≥11.57 m/s`, not a physics finding) | vector-sum model DISCONFIRMED as specified on this sample; CT-F17 gate now moot for this model form — a revised model needs its own W6-shaped test first; `E-THE-DISPLACEMENT-FILTER-ATE-THE-STRANDED-STRATUM-1` |
 | D-W2sB | α-window sweep β∈[0.85,1.15] | gated (W2s-a) | Queued | corridor α discriminator |
 | D-W7 | Corridor two-regime α field probe | gated (W6) | Queued | §10.3 physics |
 | D-CT-F17 | FRESH-sample verdict (1959–1979, N=70 candidates, V-test p<0.05 ∧ R̄≥0.35; independent adversarial spec audit MANDATORY before bars commit) | gated (W6 + audit) | Queued | the directional claim's verdict path |

diff --git a/.claude/plans/weather-w-probes-v1.md b/.claude/plans/weather-w-probes-v1.md
@@ -550,6 +550,56 @@ Checkpoint after every storm; resume skips completed `t0`.
 
 ---
 
+### RUN, 2026-08-12 (`comet_tail_w6.py` / `.json`) — B0 VOID, and the vector-sum model does NOT fit; the stranded stratum is EMPTY by construction
+
+| bar | verdict | measured |
+|---|---|---|
+| **B0 CONTROLS** | **VOID** (own pre-registered rule) | single-geo R²=**−0.104**; permuted-P_bow control R²=**−0.071**; rotated-90° control R²=**−0.062** — both controls exceed `single-geo + 0.03 = −0.074` |
+| **B1 IDENTIFIABILITY** | VOID (B0 failed) | joint R²=**−0.086** vs best-single R²=−0.104, margin +0.018 (needed +0.10) |
+| **B2 SIGN** | VOID (B0 failed) | `c_geo = +0.407` (predicted positive — **correct sign**); `c_bow = −0.0006` km⁻¹ (predicted positive — **wrong sign**) |
+| **B3 residual resultant** (descriptive) | overall: R̄=**0.153**, μ=−2.3°, p=0.646 — *below* the n=19 uniform-expectation floor (√π/2√19 ≈ 0.203); **stranded (<8 m/s): n=0**; moving (≥8 m/s): n=19, identical to overall | no clustering at all; stranded stratum **structurally empty** |
+
+**Every single-predictor model already fails on its own** — geo alone R²=−0.104, bow alone R²=−0.147, both *worse than predicting the mean*. The joint model's negligible +0.018 margin over the geo baseline is exactly what an anti-vacuity control should reject, and B0 correctly rejects it: two deliberately-wrong references (permuted, rotated 90°) score as well as or better than the real geo predictor alone. **There is no signal here for a joint fit to identify.**
+
+**No implementation bug found — checked, not assumed.** One storm was independently re-fetched and hand-audited (t0=54358); extended to the full 19-storm sample (committed in `comet_tail_w6.json`'s `B4_per_storm`, which now carries the raw `D`/`P_geo`/`P_bow`/`A_H`/`d_H`/`v_rel_ms` per storm, not only derived bearings — the first run shipped without these and the gap was self-caught before a reviewer needed to). Every value is physically sane (`A_H > 0` always, `d_H` inside the 600–2500 km annulus always, `v_rel` 2.8–27.9 m/s).
+
+> **⚠ TWO CORRECTIONS, 2026-08-12 (codex + CodeRabbit P2/Major on PR #940,
+> both real, both fixed same day) — the SIGN and the UNITS were both wrong
+> in the first draft above.** (1) **Sign:** `spine()`'s raw fit coefficient
+> points toward the storm's HIGH side (the gradient of *increasing*
+> residual pressure), while `P_geo`/`P_bow` both point toward the LOW side
+> by construction — exactly the convention `low_pole_bearing()` makes
+> explicit via its own `(ph + π) % (2π)` flip
+> (`comet_tail_f16.py:138-160`). `D` is now `−spine(...)`, matching that
+> convention. This flips `c_geo` from −0.407 to **+0.407** (now the
+> physically PREDICTED positive sign) and `c_bow` from +0.0006 to
+> **−0.0006** (now the WRONG sign, where it had looked merely near-zero
+> before) — the table above already carries the corrected values.
+> Verified algebraically AND numerically before/after the flip: R² and
+> every B0/B1 verdict are provably unchanged by this sign convention
+> (OLS is odd-symmetric in the fit target), confirmed bit-identical on
+> rerun. (2) **Units:** `D` and `P_geo` are both [Pa/km]; `P_bow` is [Pa]
+> — so `c_geo` is dimensionless but `c_bow` carries **km⁻¹**, and OLS
+> coefficients rescale inversely under column rescaling while leaving
+> R²/fitted values unchanged. **Raw `|c_bow|` was never a valid basis for
+> "no measurable weight"** — the original `|P_bow|` averaging 147× `|D|`
+> comparison compounded the same error (Pa vs Pa/km, not comparable at
+> all). The dimensionally valid measure is the FITTED CONTRIBUTION,
+> `|c_bow·P_bow|` against `|D|`, both in Pa/km:
+> **mean `|D|` = 0.745, mean `|c_geo·P_geo|` = 0.186 (25 % of `|D|`), mean
+> `|c_bow·P_bow|` = 0.068 (9 % of `|D|`)**. The geo contribution is ~2.7×
+> the bow contribution in fitted terms — modest, not "no weight" — and
+> both are consistent with the R²<0 finding that neither predictor
+> meaningfully explains `D`'s variance. **B0/B1's VOID verdicts are
+> untouched by either correction** — this changes only the supporting
+> narrative around B2, not the headline finding.
+
+**The stranded stratum is empty for a structural reason, not a physics finding.** `min(v_storm) = 12.54 m/s` across all 19 storms — comfortably above `250 km / 6 h = 11.574 m/s`, the speed CT-F14's own `displacement_km ≥ 250` qualifying filter mathematically implies as a floor. **No storm in this displacement-filtered 19-storm set can ever be "stranded" (<8 m/s) — the report §10.2 stranded-rescue reading is UNTESTABLE on this sample by construction, not refuted.** Testing it needs a sample built WITHOUT the fast-motion-selecting displacement filter (or with a filter that explicitly retains slow storms) — a design note for any future stranded-rescue probe, not a task for this one to retrofit.
+
+**What this means for the report's vector-sum model.** As specified — a global 2-parameter linear combination of a single background-high neighbor predictor and a single relative-motion bow-wave predictor, fit across 19 storms by ordinary least squares — **the model does not fit this data, and the fit is not merely weak, it is void by its own anti-vacuity control.** This does not rule out a richer version of the model (multiple neighbors, a nonlinear bow term, storm-specific coefficients) — but the specific, pre-registered, mechanistically-motivated form named in §10.2 is disconfirmed on this sample as tested. Consequence for CT-F17: its gate ("W6's result AND an independent adversarial spec audit") is now moot for the vector-sum model's CURRENT form — a fresh-sample verdict on a model that already fails its identifiability control on the STORED sample would not be a meaningful next step; the audit gate stands for any REVISED form of the model instead.
+
+---
+
 ## §4 BRIEF W2s-b — the α-field on a real H–T pair (GATED on W2s-a G2 pass)
 
 **File:** `corridor_alpha_probe.py`. Outline — finalize bars at spawn time