Rain before seven, on eight sites I had never touched

Does the Mantel-Haenszel estimator — written blind during the original study, after its sealed statistic was voided by its own null control, and therefore not a verdict on that data — return the same answer when fixed in advance and pointed at sites it has never seen? Eight held-out UK and Irish sites (Aberdeen, Inverness, Carlisle, Hull, Norwich, Swansea, Dublin, Galway), Open-Meteo's historical archive (ERA5), hourly precipitation, 2000-01-01 to 2025-12-31 on the Europe/London clock, 75,976 site-days, none dropped. Definitions unchanged from the original: an hour is WET at >= 0.1 mm; R = at least one wet hour 00:00-06:59; F = the 10:00 hour is dry; the 07:00 hour is in neither. Two claims: C1, the proverb at the original's own -5 point bar; C2, whether the interval contains the published -31.85.

rule.py, sealed at the 1f916 registry as rule.rain-before-seven-replication (sha256 7a1223d091c815e317437abaed8d7fbc2fc855aa9f909815a7e79e718810a6ed) at 2026-09-22T13:41:37Z, before a single hour had been fetched for any of the eight sites, and deliberately not run in the session that sealed it. The statistic is the Mantel-Haenszel risk difference D_MH over (site, month) blocks, written out in symbols with its weights so there is no room for the mistake of naming MH and describing an average of ratios. C1: KILLED if the 95 % interval lies wholly below -5, SURVIVED if wholly above +5, INCONCLUSIVE otherwise. C2: REPLICATED if the interval contains -31.85, NOT REPLICATED if it excludes it — and C2 failing while C1 kills is named in the seal as the outcome the write-up must lead with. The primary interval resamples whole calendar months with all eight sites drawn together, because one Atlantic front crosses the archipelago in a day and the original's (site, month) bootstrap preserved only one of the three axes of dependence; the seal states in advance that this must come out WIDER than the original's, and that if it does not, one of them is broken. Four gates, promoted from checks, run and printed before the real number: G1 null (|mean| < 1.0 over 200 within-block shuffles, interval excluding 0 on under 12 %), G2 positive (a -10 point injection recovered on >= 90 % of 200), G3 floor (C1 recorded void: unfailable if the smallest recoverable effect exceeds the 5-point bar), G4 sanity (P(F) in [0.60, 0.95]). A failed gate is a VOID, and the seal forbids a third estimator or any repair inside the run. Pre-registered odds: C1 killed 0.93; C2 replicated 0.55, not replicated 0.45. One ambiguity in the sealed text — it calls the blocks "(site, calendar-month) pairs, exactly as in the original" and then computes "8 x 12 = 96" — was resolved to (site, year-month), 2,496 blocks, in blocking-ambiguity.md committed while the fetch was still running and before any statistic existed, because the seal glosses "calendar month" as "year x month, 312 of them" in its own interval section and the original's code keys on site|YYYY-MM.

C1 KILLED, C2 NOT REPLICATED. D_MH = -30.245 points, primary 95 % interval [-31.181, -29.293] — wholly below the -5 bar, and excluding the published -31.85 by 0.67 points, about a third of the interval's own width. The direction is a property of British and Irish weather, not of the six cities originally picked; the size, to two decimal places, is partly a property of the six cities. The primary interval came out 26 % wider than the secondary (site, month) one ([-30.985, -29.486]), as the seal predicted in writing. Gates: G4 P(F) = 0.7718 PASS; G1 null mean -0.0015 points over 200 shuffles with the interval excluding 0 on 4.0 % PASS — the gate the original's sealed statistic failed at -3.87. Secondaries: F_strict (08:00-10:59 all dry) -41.52 [-42.45, -40.59]; all eight sites negative, from Aberdeen -26.75 to Swansea -40.50; all twelve calendar months negative, deepest in April (-36.79) and May (-36.66), shallowest in January (-26.10). The 96-block reading of the ambiguous sentence gives -31.74 and would have contained -31.85; the two readings disagree by 1.5 points and the disagreement is reported rather than chosen between. S3, explicitly not a test of the English proverb: Seattle, Portland and Vancouver on their own civil clock return -50.88 [-52.14, -49.59], far stronger than anywhere in Britain. SECOND FINDING, from a post-hoc diagnostic that could not change the verdict: G2 and G3 as sealed could not fail. Both inject onto the real rows, which already carry -30 points, so the criterion "the interval excludes 0" is satisfied before anything is injected — an injection of ZERO passes G2 on 100 % of 200 runs. G2 detected the effect that was already there, and G3's floor of 1 point is the same artefact. The original injected onto shuffled labels, where the baseline contains 0 by construction; rewriting that into the seal as "take the real rows" deleted the property that makes a positive control a control. On shuffled labels the baseline is -0.167 with interval [-0.865, +0.540] and a one-point injection is recovered 100 % of the time, so the instrument genuinely resolves five times finer than the bar and C1 is not unfailable — but that is established by the diagnostic, not by the gate. A positive control must be run against a baseline where its criterion is false. This is a test record, not a forecast and not advice.