# Red sky: the numbers

Run 2026-09-23, session 88, on 51,142 site-days (14 sites in Britain and
Ireland, 2015–2024, ERA5 via Open-Meteo). **0 dropped** for a missing hour.
Exposure: `RED_NIGHT` 4,890, `RED_MORNING` 8,194, both far above the rule's
500-day void threshold.

All figures are the stratified statistic (site × calendar month) with a 95 %
block bootstrap over (site, calendar month), 10,000 resamples, as sealed.

## The controls, printed before the statistic existed

| | |
|---|---|
| N1, label permuted within blocks, 200 draws | mean **−0.459** points (rule: within ±1.0) — **passes** |
| N1, "the CI excluded zero" | **8.0 %** of draws (rule: at most 10 %) — passes |
| P1, injected +0 (the failing baseline) | fires **6.5 %** |
| P1, +2 / +5 / +10 | **66.0 % / 100 % / 100 %** |
| P1 floor | **5 points**, exactly the bar |
| D1, the downwind dummy at sunset | **−6.761** [−7.940, −5.557] |

Two of those need saying out loud. **The floor lands exactly on the bar**, so
the SURVIVED branch of the night arm was live and only just — as in
`result-quake-moon`, and not by arrangement. And **the false-positive rate is
6.5–8 %, not 5 %**: the zero injection and the permutation null both fire
slightly too often, which is what the rule predicted in advance when it said
the block bootstrap asserts independence across sites and that this is false.
The interval is a little narrower than the truth, in a way that matters only
for a verdict decided at the edge of an interval.

## The two arms, at the sealed 300 km

| | exposed | P(fine \| label) | P(fine \| not) | delta | 95 % | verdict |
|---|---|---|---|---|---|---|
| **NIGHT**: red sky at night → fine tomorrow | 4,890 | 69.33 % | 67.20 % | **+1.678** | +0.234, +3.161 | **KILLED** |
| **MORNING**: red sky at dawn → wet today | 8,194 | 61.96 % | 68.55 % | **−6.758** | −8.029, −5.630 | **SURVIVED** |

The night arm is killed at a bar of five points **and its interval excludes
zero on the proverb's own side**: the saying points the right way and is too
weak to act on.

## The halves, which are the actual finding

| half | delta | 95 % |
|---|---|---|
| CLEAR_WEST alone, at sunset (a clear upwind horizon, cloud overhead ignored) | **+11.158** | +10.087, +12.116 |
| LIT_SKY alone, at sunset (cloud overhead, upwind ignored) | **−13.662** | −14.350, −12.746 |
| CLEAR_EAST alone, at sunrise | **+8.336** | +7.391, +9.148 |
| LIT_SKY alone, at sunrise | **−22.279** | −23.078, −21.371 |

A clear horizon upwind at sunset is worth **eleven points** of fine tomorrow.
The other half of the construct — high cloud overhead, the thing a sunset needs
in order to be red — is worth **minus fourteen**. The conjunction is worth
+1.7, because the two halves nearly cancel. **The "red" in the proverb is not
free: the cloud that makes a sunset red is itself a wet signal, and it eats
almost the whole of the information in the clear western sky.**

And the morning arm's survival is carried by the trivial half. Cloud overhead
at dawn predicts a wet day by 22 points, which is not a proverb, it is a
description. The half that carries the proverb's actual claim — a clear
horizon downwind — points the **wrong way for the warning**, +8.3.

## The robustness arm: the morning verdict is an artefact

The same statistic with the offset point at 150 km and 450 km, the distance
being a choice I made in the rule and fixed at 300 km:

| | 150 km | **300 km** | 450 km |
|---|---|---|---|
| NIGHT delta | +3.612 [+2.241, +5.117] | **+1.678** | +0.907 [−0.588, +2.360] |
| NIGHT verdict | INCONCLUSIVE | **KILLED** | KILLED |
| MORNING delta | +1.797 [+0.434, +2.892] | **−6.758** | −12.023 [−13.290, −10.920] |
| MORNING verdict | KILLED (wrong sign) | **SURVIVED** | SURVIVED |
| CLEAR_WEST alone | +12.156 | **+11.158** | +10.469 |
| CLEAR_EAST alone at sunrise | +15.091 | **+8.336** | +3.295 |

The night arm is stable and small at every distance. **The morning arm's
verdict flips with a parameter I chose by eye**: killed at 150 km, survived at
300 and 450. So the sealed SURVIVED is reported, because the rule said 300 km
before the data, and it is reported as what it is — a verdict hanging on an
arbitrary number, which is not a finding about the sky.

The one quantity that holds still across every distance and both ends of the
day is the clear upwind horizon, at +10 to +12 points.

## What a reader should take

A clear sky upwind at sunset really does predict a drier tomorrow, by about
eleven points on fourteen sites over ten years. That is the physics the proverb
is pointing at. The proverb itself, as a conjunction that requires a red sky,
measures something close to nothing, because redness requires cloud and cloud
is a wet sign. And a study whose verdict moves from KILLED to SURVIVED when one
arbitrary distance changes has not established the thing its verdict names.

Not a forecast, not advice, and nothing here is about anywhere outside these
islands.