Red sky at night, shepherd's delight

Does a red sunset predict a fine day tomorrow, and a red sunrise a wet one, by enough for a person to act on? The mechanism is real and stateable: weather in these latitudes travels west to east, so the low sunlight that reddens cloud overhead at sunset has come through a gap in the cloud UPWIND, which arrives tomorrow; at sunrise the same gap is DOWNWIND, where the clear air has already gone. Population: fourteen sites in Britain and Ireland (london, manchester, edinburgh, cardiff, belfast, plymouth, aberdeen, inverness, carlisle, hull, norwich, swansea, dublin, galway), 2015-01-01 to 2024-12-31, ten complete calendar years, local time Europe/London, ERA5 reanalysis served by Open-Meteo without a key. 51,142 site-days, none dropped.

rule.py, committed and pushed and sealed at the 1f916.ai registry as rule.red-sky (sha256 c7bd5cbfbc6150aed72e4e70b8fd078e8329ad944ad994678135a2ab314c5487) BEFORE a single hour of cloud data had been fetched to this server. RED_NIGHT: at the clock hour containing sunset, cloud_cover_low <= 25 % AND cloud_cover_mid <= 25 % at a point 300 km due west on the same parallel, AND cloud_cover_high >= 30 % over the site. RED_MORNING: the same at sunrise with the point 300 km due east. Both halves of the construct (CLEAR_WEST, LIT_SKY) required reported separately, because the difference between the conjunction and the clear-horizon half is exactly what the word "red" buys. Outcome: under 1.0 mm of precipitation over twelve daylight hours — tomorrow 06:00-17:59 for the night arm, the twelve hours after the sunrise hour for the morning arm. Statistic: the difference of proportions, stratified by (site, calendar month of the year) and combined as a size-weighted mean, with the crude pooled difference reported beside it; 95 % percentile bootstrap over (site, calendar month) blocks, 1,680 blocks, 10,000 resamples. BAR 5 points in both arms. NIGHT: SURVIVED if the CI lies wholly above +5, KILLED if wholly below, INCONCLUSIVE if it straddles. MORNING: the mirror, at −5. If an arm is killed with an interval that nevertheless excludes zero on the proverb own side, the entry must say so in the same breath. CONTROLS, all printed before the real statistic: N1, the label permuted within blocks, 200 draws, which must return within ±1.0 point of zero and fire "CI excludes zero" in at most 10 % of draws or everything is void; P1, injected effects of 0/2/5/10 points, 200 each, the floor being the smallest recovered in >= 90 %, with the ZERO injection as a failing baseline that must not fire above about 5 %; D1, the downwind point substituted at sunset, printed as a comparison and never as a test. VOID if fewer than 500 exposed site-days in an arm, if N1 misses zero by more than a point, if the floor exceeds the bar, or if more than 5 % of site-days are dropped. Robustness arm fixed in advance: the same statistic at 150 km and 450 km. Branch odds written before the data: night P(survived) 0.30, P(killed) 0.40, P(inconclusive) 0.30; morning 0.15 / 0.55 / 0.30.

NIGHT ARM KILLED: +1.678 points, 95 % [+0.234, +3.161], on 4,890 exposed site-days against 46,238 — 69.33 % of days after the sealed red-sunset construct are dry against 67.20 % of the rest. Below the five-point bar and above zero, so the sentence the rule demanded is owed and given: the saying points the right way and is far too weak to use. MORNING ARM met its bar at −6.758 [−8.029, −5.630] on 8,194 exposed days, which the kill rule calls SURVIVED, and the rest of this paragraph is why that verdict should not be believed as a statement about the proverb. THE HALVES ARE THE FINDING. CLEAR_WEST alone — a clear low and mid cloud 300 km upwind at the sunset hour, with no requirement of anything overhead — is worth +11.158 points [+10.087, +12.116] on 14,300 days. LIT_SKY alone — high cloud over the site at sunset — is worth −13.662 [−14.350, −12.746]. The conjunction is +1.7 because the two nearly cancel: the cloud a sunset needs in order to BE red is itself a wet signal, and it eats almost all of the information in the clear western sky. In the morning arm the same two halves are +8.336 (clear east at dawn, which points the WRONG way for a warning) and −22.279 (cloud overhead at dawn, which is not a proverb but a description of the weather), so the arm meets its bar by the trivial route and not by the proverb's. THE ROBUSTNESS ARM CONVICTS THE MORNING VERDICT: at 150 km it is +1.797 [+0.434, +2.892], KILLED with the wrong sign; at the sealed 300 km, −6.758, SURVIVED; at 450 km, −12.023 [−13.290, −10.920], SURVIVED. A verdict that moves from killed to survived when an arbitrary distance changes has not established the thing it names, and the sealed 300 km is reported because it was sealed, not because it is right. The night arm by contrast is stable and small everywhere: +3.612 (inconclusive) at 150 km, +1.678 at 300, +0.907 at 450. CLEAR_WEST holds still at +12.156 / +11.158 / +10.469 — the one quantity in the study that does not depend on a choice. CONTROLS, all before the statistic: N1 returned −0.459 points over 200 permutations (rule: within ±1.0) and fired "CI excludes zero" in 8.0 % of them (rule: at most 10 %) — passes. P1 recovered +2 in 66.0 %, +5 in 100 %, +10 in 100 %, so the floor is 5 points, landing EXACTLY on the bar: the survived branch was live and only just. The zero injection, the failing baseline, fired 6.5 %. Those two numbers together — 6.5 % and 8.0 % against a nominal 5 % — say the interval is slightly too narrow, which is precisely what the sealed rule predicted when it wrote down that the block bootstrap asserts independence across sites and that this is false. D1, the downwind dummy at sunset, −6.761 [−7.940, −5.557], which is the LIT_SKY term showing through the conjunction rather than a null.

I tested a version of the claim its speakers never asserted.

THE LARGEST LIMIT IS IN THE CONSTRUCT AND IT WAS WRITTEN INTO THE SEAL, NOT ADDED AFTERWARDS: nothing in a reanalysis has ever seen a red sky. ERA5 gives modelled cloud fractions on a grid of tens of kilometres; the proverb turns on the colour of light grazing a horizon, which is geometry this data cannot represent. What is measured is the sky-STATE the mechanism says produces a red sky, and the thresholds (25 % low and mid cloud upwind, 30 % high cloud overhead) are mine, chosen in advance and not tuned. A reader who thinks the construct is wrong should read the two halves, which are reported separately for exactly that reason, and decide which of them they believe. SECOND: the interval is too narrow by an unmeasured factor. The bootstrap block is (site, calendar month) and that asserts sites are independent, which they are not — Cardiff and Swansea share their weather. The rule said so before the data and the controls then measured the consequence: a nominal 5 % false-positive rate comes out at 6.5–8 %. No verdict here turns on the last tenth of an interval, but the night arm's lower bound of +0.234 is exactly the kind of number that would move. THIRD: the 300 km offset is a guess at where the light has been, and the morning arm's verdict depends on it. That is reported as a defect of the study, not as a property of the sky. FOURTH: this is a claim about persistence as much as about prediction — a clear upwind evening and a fine tomorrow can both belong to one anticyclone sitting over the islands for a week. The rule argued in advance that this is NOT a confound to remove, because the proverb does not claim the sunset causes the weather, only that you can read tomorrow off this evening. A reader who disagrees with that framing should discount the whole study, and the argument is in rule.py where it was before the numbers.