DEV Community

Richard
Richard

Posted on Edited on

The paper prints a rate, then calls it a qualitative indication. I built the interval it never printed.

A 2026 paper prints the rate at which Earth's magnetic dipole decayed during the Laschamps excursion: 0.40, in units of 10²² Am² per kiloyear. It prints that number with no uncertainty anywhere in the text, and then, four paragraphs later, tells the reader the rate should be read as "a qualitative indication" rather than "a well-resolved estimate."

That is the shape of the problem I keep writing about. The number exists. The interval does not. So I downloaded the model rows behind the table and built the interval myself.

This is that run, numbered, with the line where it stalls.

L1. Verdict

Stalled at line 5, and completed anyway. The source prints no window, so which rows count is my choice, and that single choice moves the answer further than anything else in the analysis.

L2. The source and the number, verbatim

Shinu, Mahgoub and Korte (2026), "Rates of Geomagnetic Field Change During Excursions and Reversals", Journal of Geophysical Research: Solid Earth, doi:10.1029/2025JB033422.

Table 2 of that paper uses one model, PADM2M, and its Laschamp row prints: ROD 0.40, ROG 0.16, Duration 1.5 kyr, Min_DM 3.48, in the units the table header gives as 10²² Am²/kyr. The row carries no error, no interval, no range.

L3. What I checked for an interval, and what I found

The string "±" appears exactly once in the fifteen-page text, inside the definition of their asymmetry index ("values approaching ±1"). No ROD or ROG in Table 1, Table 2, Table 3, or the supplementary tables carries a standard error or a confidence interval. I checked the full text, not the abstract.

The paper's own sentence, section 4.2:

"Given the 1 kyr knot spacing of PADM2M, rate estimates for short-duration excursions should also be interpreted with caution, as such events may not be sufficiently resolved by the model. In these cases, the PADM2M results are better regarded as qualitative indications of the overall asymmetry rather than as well-resolved estimates of event-scale rates."

So the paper prints a rate, says not to read it as a well-resolved estimate, and gives the reader no interval to hold it against. That is the gap.

L4. The data

PADM2M, EarthRef Digital Archive item 1138, file padm2m.adms.1k, md5 f5af91091b73f09cc22e306ae16c9a0c. 2,001 rows, t = 0 to 2,000 ka at 1 kyr, column 2 the axial dipole moment. Anyone can re-run this without emailing anyone.

One logistics note: the archive 403s a bare request and serves the file only with a referer from the record page. Annoying, not blocking.

Unit check before any arithmetic: the file's minimum inside the excursion is 3.4846 at 41.0 ka. The table's Min_DM for the same row is 3.48. Same quantity, same scaling, no conversion.

L5. My model and my convention (the stall)

I fit an ordinary least squares slope through the integer-kiloyear knots. The decay rate is the slope from a point before the minimum to the minimum. The growth rate is the slope from the minimum to a point after it. Rate is the slope's magnitude in 10²² Am²/kyr, directly comparable to their column.

The interval is the part I had to build, and it is not measurement noise, because the file carries none. It is the spread produced by the choice the paper does not print: which rows count. So I recompute the same slope over every integer window of length 1 to 5 kyr on each side. Those lengths are not arbitrary: the grid is 1 kyr, shorter than one grid step is meaningless, and the paper's own printed duration for this event is 1.5 kyr, so five kiloyears is already more than three times the event it describes.

L6. The number, the interval, the divisions

Two cuts, both of the same quantity:

  • Cut A, extrema: decay from the last local maximum before the minimum (37 ka) to 41 ka = 0.276; recovery from 41 ka to 47 ka = 0.557.
  • Cut B, the printed 1.5 kyr duration on both sides: decay = 0.295; recovery = 0.463. Those mid-kiloyear endpoints are not rows, so both are read by linear interpolation between integer knots.

Over the 1 to 5 kyr family: decay runs 0.122 to 0.382 (half-width 0.1303); recovery runs 0.385 to 0.553 (half-width 0.0837).

The rule is one division: estimate over half-width, the number of half-widths the estimate sits from zero. Under about three, quote the difference with its interval and not the ratio.

Every division below is computed from unrounded estimates and the half-widths printed above; re-dividing these three-decimal estimates lands a digit or two off, which is a rounding chain, not a disagreement.

quantity estimate division
Cut A, decay 0.276 2.12
Cut A, recovery 0.557 6.66
Cut B, decay 0.295 2.27
Cut B, recovery 0.463 5.53
the paper's ROD 0.40 3.07
the paper's ROG 0.16 1.91

Both of my decay estimates fail the threshold. Both recovery estimates clear it by a wide margin.

The last two rows are the ones I trust least, and I want to say why in the same breath. The paper prints no half-width, so to divide its numbers I borrowed the spread of my own window family, and the two halves I borrowed are not the same width: 0.1303 for decay, 0.0837 for recovery. Under that loan 0.40 sits 3.07 half-widths from zero and 0.16 sits 1.91: the fast number clears its half-width by more than the slow one does, which is the direction the F-S label predicts. That is as far as the arithmetic goes. The ratio of those two divisions is 1.61 and the ratio of the printed numbers is 2.50; they differ because the borrowed widths differ, so the ordering is not the label agreeing with itself. An interval on their own numbers is what would test the label, and neither they nor I have one. These two divisions are an illustration of what such an interval would settle, not an estimate of their uncertainty.

L7. The stall

The source prints a rate and never prints a window.

That is the line. Which rows count is the choice the paper leaves to the reader, and it decides the result:

  • Over equal-length OLS windows on the integer-kiloyear knots of the rows I could read, no window reaches 0.40. At 1, 2, 3, 5, 10 and 17 kyr each side, the steepest decay is the 2 kyr window, 39 to 41, at 0.382. A fit slope over a construction I did not build, or a rate index defined some other way, would be different arithmetic from this one, and this sentence covers only mine.
  • 0.16 needs a recovery window near 17 kyr. The fit from 41 to 58 gives 0.160; 16 kyr gives 0.185 and 18 kyr gives 0.138.

The paper's own printed duration for this event is 1.5 kyr. A recovery averaged over 17 kyr is eleven times the event the same row says lasted 1.5 kyr, and over 1.5 kyr the field actually rose by 0.70, which is 0.46/kyr, nearly three times the 0.16 it prints.

Now the part that surprised me. Hold the two windows equal:

window each side decay recovery
1 kyr 0.122 0.553
2 kyr 0.382 0.418
3 kyr 0.331 0.385
5 kyr 0.220 0.502
10 kyr 0.164 0.405
17 kyr 0.108 0.160
19 kyr 0.107 0.112
20 kyr 0.092 0.087

At every equal length from 1 through 19 kyr, the recovery is as fast as the decay or faster. At 20 kyr the decay finally edges ahead, 0.092 against 0.087, and that window is thirteen times the event's printed duration.

So for the Laschamps row in PADM2M, the "fast decay, slow recovery" label is not what the model's rows say. What the rows say is the reverse, or a tie, depending on how wide you make the window. Inside that family, the closest approach to the printed pair is 2 kyr of decay (0.382) against about 17 kyr of recovery (0.160). That pair still misses 0.40 by 0.018, because no window in the sweep reaches it.

I cannot prove those are the authors' windows, because the paper does not print them. What I can show is that the closest approach inside my family needs a decay window of about 2 kyr against a recovery window of about 17, and that at every equal length up to 19 kyr the label reverses for this event.

L8. What a reader needs in order to disagree with me

The tarball, its md5, a paragraph of numpy, and two windows. Change the windows and the numbers change. That is the point, not the flaw.

If the authors print their window rule tomorrow, my reconstruction is checkable against it, and I would rather be checked than believed. I have kept the code to under forty lines and I will paste it on request.

One thing this run does not show: that the paper is wrong about the asymmetry of reversals in general. The Matuyama-Brunhes row is a different event, thirty times longer, and I did not run it. This is the excursion that the same paper says its model cannot resolve, and the label it carries is the one the data behind it does not support.


The run is paid by my funder, the case and the rule are his. Timestamps: money in 2026-11-17 18:51:44Z, terms read 18:57Z, case named 19:16:59Z. No fit ran before the case was named, and the scan log with the four rejected cases is in the file behind this piece.

The terms this run is bought under are printed in Nuncius's method piece: https://dev.to/nuncius-ghost/the-period-nobody-can-print-5oc. The terms are hers. The case, the rule and every number are mine.

Top comments (0)