For months this site told you, town by town, the historical odds of a clear sky on 12 August. It was never a forecast: it was climatology — thirty 12ths of August of ERA5 reanalysis between 1996 and 2025, folded into the scoring model at a 30% weight — and that number had an awkward property: it can only be graded afterwards. The eclipse has happened and the ERA5 archive now holds that day, so there is exactly one honest thing left to do: download the real cloud cover and put it next to ours.
What was downloaded, and at what hour
The same 404 ERA5 cells of 0.25° that covered the totality band, for the single day 2026-08-12: daily mean cloud cover, temperatures, precipitation and — the part that matters — hourly cloud cover, both total and low cloud. Eleven requests to the Open-Meteo archive, and all 404 cells answered. Nothing is missing.
The hour matters more here than in almost any other eclipse. This one happened at sunset: maximum fell between 20:27:18 and 20:33:30 local time depending on the town, so a daily average describes what nobody saw. For each cell we average the two ERA5 hourly samples that bracket that instant (20:00 and 21:00) — the same rule the site already uses for the 2027 eclipse window. One resolution warning: each town inherits its cell’s observation, ~25 km across, exactly as it inherited the climatology. This is not a measurement taken in the village square.
One thing was deliberately not downloaded: sunshine hours. ERA5 does not model the eclipse itself, so its radiation for that day does not reflect the real darkening, and quoting it would be inventing a number that looks like a measurement.
The headline: 46% expected, 87% real
Averaging the probability we gave to all 3,963 towns in the band, we expected 46.1% of them to be clear — using the project’s canonical threshold, cloud cover below 30%. At the hour of maximum, the real figure was 86.9%; measured by daily mean instead, in case anyone suspects the choice of hour, 88.5%. Mean cloud cover across the whole band at maximum was 12.9%, and only 2.4% of towns sat under genuinely overcast sky.
12 August 2026 came out, in short, far better than it had any right to. And that, stated precisely, does not mean the prediction failed: a 46% probability is not refuted by one good day, any more than a die is loaded because it rolled a six. What a single day can test is something harder and more useful — whether the map ranked places correctly.
Where we were right: the Cantabrian coast
At the bad end, yes — though not cleanly. The only two provinces that genuinely lost the eclipse to cloud were Cantabria, with 27.5% of its towns clear, and Asturias, with 28.2%: the two worst observed figures among the 31 provinces with towns in totality. Both sat at the bottom of our own ranking — we gave Asturias 12.1%, the worst bet in the entire band, and Cantabria 18.1%. The map pointed at the disaster where the disaster happened.
What it missed is that the bottom of that ranking was not one block. A Coruña (16.0% expected) and Lugo (17.4%) were rated almost as badly as Asturias and finished with 86.4% and 69.7% of their towns clear. Our climatology treated “the wet north” as a single thing; on the day, the north split down the middle.
The detail that clinches the half it got right is the kind of cloud. Along the coast between Cantabria and Lugo there were cells with 100% cloud cover and 100% low cloud — not high cirrus, but the marine stratus that eats the western horizon. With the Sun between 2° and 12° up, that is precisely the cloud that kills a sunset eclipse. Mean low cloud in Cantabria at the hour of maximum was 47.7%, against 2.7% for the band as a whole.
That same figure rescues a good share of the towns that failed the cut. Of the 71 cells that came in above 30% cloud, 45 had the cloud sitting high and the horizon clean. Only 26 cells out of 404 carried enough low cloud to ruin the view.
Where we were wrong: the Mediterranean, our own favourite
The good end of the ranking is where the model missed, and not narrowly. The Mediterranean coast was the jewel of our climatology — our piece on thirty years of clouds pointed straight at it — and on the day it finished below the band average. València held the best figure in Spain (68.5% expected) and observed 67.4%: dead on, and yet next to last on a day when the interior plateau went almost perfect. Castelló, forecast at 59.7%, landed on 61.5%. The biggest outright error was Lleida: 54.0% expected, 42.3% observed, the only province to fall by double digits. In the map’s defence, the other half of that recommendation held: Ibiza and the rest of the Balearics came in at 92.5% of towns clear.
Symmetrically, the provinces we wrote off ran the table. Bizkaia went from 18.7% expected to 100% observed, Navarra from 30.1% to 100%, La Rioja from 37.6% to 100%. Fifteen of the band’s 31 provinces had every single town clear. Anyone who followed the map down to Castelló did see the eclipse — but through more cloud than someone who stayed in Pamplona out of sheer inertia.
How much skill the model had, in numbers
To avoid arguing this with adjectives, two metrics computed over the 404 cells — not over the 3,963 towns, which share cells and are not independent samples:
- Brier score of our climatology: 0.2919. A dumb prediction assigning the band’s own average rate everywhere scores 0.2944. The improvement is 0.008 — effectively none.
- AUC (ranking ability, immune to the day being better than normal): 0.606. Pick two cells at random, one clear and one not, and our number picked the right one about 6 times in 10. Better than a coin, far worse than a map should be.
Calibration says the same thing in more detail. Of the 72 worst-rated cells (14.5% on average) 54.2% came out clear; of the 68 rated at 46.2%, 97.1% did; and of the six best-rated cells of all, at 71.8% expected, only half. Everything shifted upward, and the ordering broke right at the top.
The top 20, checked against reality
Of the twenty highest-scoring towns in the current ranking, 19 had clear sky at the hour of maximum. The one that didn’t, Munébrega (Zaragoza), came in at 31.5% — a point and a half over the cut, with low cloud at 0%, so it almost certainly saw the eclipse anyway. Torreblanca, the number one, was the tightest of those that did pass, at 23.5%. The ten towns from Burgos and Palencia on the list recorded a flat 0%.
A necessary caveat: that top 20 is not identical to the one we published in the ranking piece. Neither list is a typo, and the reason is the last piece of this retrospective.
The correction we made mid-campaign
On 25 July an internal audit found that the horizon scoring was using the Sun’s altitude exactly as the IGN publishes it — rounded to whole degrees. With the Sun 2-10° above the horizon, half a degree of error is a full solar radius: enough to flip the “Sun visible / Sun blocked” verdict in towns with a tight margin. It was replaced by the refracted topocentric altitude computed for each town’s own moment of maximum.
The effect: 3,564 of the 3,963 towns changed score, sixteen flipped their verdict — ten from visible to blocked and six the other way — and the count of towns denied totality by their own terrain rose from 85 to 89. The article we published about them is still called “85 towns inside the band that won’t see totality” and keeps its URL: breaking it would break the links of everyone who cited us. The correct figure today, and the one the rest of the site uses, is 89.
Which brings the day’s final irony. Of those 89 towns whose own terrain stood between them and the Sun, 66 — three in four — had a clear sky on 12 August. Their mean cloud cover was 20.4%, better than Spain’s average. Between them they missed a totality that would have lasted 83.4 seconds on average, under a perfect sky, because of a hill.
What this data cannot say
One day is a data point, not a series. These figures describe 12 August 2026 and cannot establish whether the site’s climatology is good or bad in general: with 404 strongly correlated cells and a single date, any confidence interval would be decoration. Nor are they field measurements — a cloud parked over one valley for ninety seconds does not fit in a 25 km grid box.
What does survive, and carries straight into the total eclipse of 2 August 2027, is the methodological lesson: climatology is a tool for ruling places out, not for picking one. Picking one needs the actual forecast from the three days before, and no map published a year in advance substitutes for that.