Ground truth · live from NSIDC, recomputed in your browser

The Record Is Five Days Wide

Every story about Arctic sea ice quotes the same number: 3.39 million square kilometres, 17 September 2012, the lowest the Arctic has ever been. That number is an average of five days. The lowest day actually recorded is a different number on a different date, and NSIDC's own public API will hand you either one, depending on a single field in the query.

The Sea Ice Index is the record. It has run since 26 October 1978 and it is where the minimum you have read about comes from. Its published daily graph, and the sentence in every press release, use a five-day trailing mean: today's value is the average of today and the four days before it. NSIDC says so plainly in the user guide, and says why:

The daily extent series is smoothed using a 5-day trailing mean. The extent value for a given day is averaged with the extent value from the previous four days, and this average value is plotted. NSIDC, Sea Ice Index (G02135) Version 4 user guide

So the famous minimum is not a day. It is a window five days wide, dragged across the melt season until it finds its lowest resting place. Below is NSIDC's request for the 2012 season, exactly as this page issues it. One field is a button.

not yet

Before it tells you anything new, it reproduces something published

A page that recomputes a moving number can say anything, because by the time you check it, it has moved. So this page earns the right to report tonight's figure by first reproducing figures NSIDC printed on a fixed date, using the same code path, from the primitive daily series alone. The trailing mean below is computed in your browser out of /data/{year}, which is the unsmoothed series. NSIDC's own smoothed endpoint is not used to produce our column.

The two minima are the ones that matter, because they are the ones the world repeats. NSIDC's post of 17 September 2025 says the 2025 minimum was 4.60 million square kilometers, reached on 10 September 2025, and gives the satellite-era record as 3.39 million square kilometers on 17 September 2012. Both are five-day means, and both fall out of our own arithmetic on the raw dailies, on the printed digits and the printed dates.

What this page does not do, stated plainly. An earlier version of this instrument reproduced these figures from the 25 km satellite concentration grids, one level below NSIDC's own arithmetic. Those grids sit on a host whose robots.txt reads Disallow: /, so we do not fetch them, and no number here descends below the daily series NSIDC publishes. This page checks NSIDC's derived numbers against NSIDC's primitive numbers. That is a real check and it is one level shallower than the one we wanted, and you should hold it to exactly that.

A rounding difference, then. Who cares.

Here is who cares. Rank every complete year in the satellite record by its annual minimum, once on the series NSIDC publishes as data and once on the series NSIDC quotes in prose. The two orderings are not the same ordering.

The reason is not subtle once you see it, and NSIDC states it in the same user guide, two paragraphs after describing the smoothing:

the average value on the last day will be higher than the true daily value when ice is decreasing (during the melt season) and lower than the true daily value when ice is increasing (during freeze-up). The true, unaveraged extent values are available in the daily extent data files. NSIDC, Sea Ice Index (G02135) Version 4 user guide

A trailing window looking backwards during a melt is always carrying four days of ice that has already gone. So the smoothed minimum is higher than the true one, in every single one of the 47 years, without exception. That is not an error. It is what a trailing mean is for. But a year whose melt ended in a sharp, short trough loses more to the smoothing than a year that bottomed out slowly, and when several years sit within a few hundredths of each other, that difference is the whole ranking.

The date is a different story, and it is worth being exact about, because the tidy version of this sentence is wrong. The smoothed minimum usually arrives later, on average 1.74 days later. It does not always: in three of the 47 years it arrives earlier instead. Separately, the widest disagreement of all is 1997, where the smoothed minimum lands 20 days later than the unsmoothed one. That is the ordinary direction taken to an extreme, not a fourth exception. The cause of both is visible in the chart above if you pick those seasons: a melt with two nearly equal troughs. The single lowest day belongs to one of them, and the five-day mean can settle on the other. The years are listed here rather than rounded off, because a rule with three exceptions is a more useful thing to know than a rule.

How big is a flip, though?

"Sixteen of forty-seven change rank" is a count, and a count with no magnitude beside it is the easiest way to make a small thing sound large. So here is the magnitude. Those sixteen moving years are produced by a handful of pairs of years swapping over, and each swap rests on the distance between that pair on the unsmoothed series. Every one of those distances is printed below, none of them rounded up.

The smoothing is not decoration. It was built to cover a real hole.

It would be easy, and wrong, to read all this as NSIDC dressing up its numbers. The opposite is true, and the evidence is in the data itself. Here is how many days of the 92 between 1 August and 31 October carry an observation, year by year.

The early record is not worse ice data. It is a different satellite duty cycle: SMMR, the first instrument in the chain, sampled every other day. A daily series with every second day missing is not a daily series, and a five-day trailing mean over a gap-filled version of it is a perfectly reasonable thing to publish. The smoothing was built for that hole, the hole closed in the late 1980s, and the smoothing stayed. That is the whole story, and it is a normal one.

The other field in the same URL

One path segment along, extent becomes area. Extent counts every grid cell that is at least 15 per cent ice, and counts the whole cell. Area counts only the ice. On the lowest day the Arctic has ever had, the gap between them is this:

Neither is the wrong number. Extent is the one with the long, consistent record and the one every headline uses. But when you read that the Arctic fell to 3.34 million square kilometres, roughly a third of what was counted was open water inside the ice edge. Not quite all of the gap is open water: 0.029 of it is the circle at the pole the sensor could not see, which extent assumes is ice and area leaves out, and the figure in the panel above has that subtracted. It survives the subtraction and stays about a third. NSIDC's user guide is direct about the 15 per cent cutoff too: values below it "are too uncertain to use". We have no instrument here that could move that cutoff and show you the consequence, because the concentration grids it would need are on a host we do not fetch. It is a choice, it is theirs, and this page can only tell you it exists.

Now tonight's number, which is why this page is worth reloading

HTTP 200 is not freshness. A server can answer, cheerfully, with well-formed JSON whose newest row is two years old, and this wave has already caught two feeds doing exactly that. So the check above is not on the status code and not on the Last-Modified header. It is on the newest date inside the data, and it is not decorative: past six days of lag this page stops printing a current-season number at all rather than printing one with a caveat under it. As it happens there is no header here to disagree with. This service sends no Last-Modified on any request anyone has made of it, which is also why the "has it moved since you were here?" line above is answered from the newest row in the data rather than from the source's clock. That is the signal this page actually has, so it is the one it uses.

A claim you can score against us, in about five weeks

This is a climatological extrapolation and nothing more. No atmosphere, no ice motion, no model: just the ratio between the mid-August value and the eventual minimum in each of the last fifteen years, and the spread of those ratios. Its stated hazard is that the ratio was fitted mostly on SSMIS-era years and is being applied to an AMSR2-era August, and 2025, the only complete AMSR2 year, sat near the top of that distribution. If it misses, the interval and the method are on the page, so the miss is informative.

Where this instrument stops saying anything

The window buttons go past 5. They should, because the place this page's comparison dies is worth walking to. Push the window to three weeks and every year's minimum gets dragged up toward its own seasonal mean, until what you are reading is not a melt season at all: it is a low-pass filter, and you are watching the filter.

A correction, kept in place of what it replaced. An earlier version of this section finished that thought with a mechanism: that as the window widens the differences between years shrink faster than the ranking can respond, so the ordering settles down and starts to look like agreement. It reads well, nothing on this page measured it, and it is false. The measurement is the chart below, and it says the ordering does the opposite of settling. The sentence is gone rather than softened, because it was not too strong, it was wrong.

measuring

Two settings on that axis carry published meaning and the rest are illustration. 0 is the series NSIDC publishes as data. 5 is the series NSIDC quotes in prose. Everything else is a number no one has ever cited, and the page labels it that way while you are looking at it, rather than in a footnote afterwards. There is a gap at 1: that value returns HTTP 500 from the service. It is a bug in their API, it is not meaningful, and we are not going to dress it up as one.

The ledger

Every verifier run appends one line to research/five-days-wide/record.jsonl, committed. It stores the newest data date, the running minimum on both series, our five-day mean beside NSIDC's with the signed difference, any date whose value changed since the previous run, and the live status of the forward claim. The first line and the newest line, side by side: