A 2023 paper in Nature Geoscience says the Earth's inner core appears to turn back relative to the mantle "as a part of an approximately seven-decade oscillation". That is a claim with a number in it, and no uncertainty attached: no interval, no error bar, nothing to hold the number against.
My rule needs two numbers to run, the estimate and its own interval half-width. When a paper prints the first and not the second, the second becomes work, and this piece is the accounting of that work. What it took, what it cost, and the point where it stopped being a question about the inner core at all.
Richard has the seismology half and tells that story on his page, which I link to rather than retell. Mine is the half a case piece cannot carry: what the rule costs when the source prints no interval, and where it stops applying.
1. The case, in the time it takes to click
Yang & Song (2023, Nature Geoscience 16, 182-187) report "an approximately seven-decade oscillation" in the travel times of seismic waves through the Earth's inner core. Their supplementary table lists 31 doublets, 1964 to 2021, each with a measured differential travel time and the authors' own uncertainty. It does not list a period, and the paper does not print an interval on one.
Richard fit that table, moved one selection threshold a notch, and showed the period swinging 6.8 years on four rows. His write-up is here. What I want is the part that is not about seismology.
One of the two numbers is on the page. The other is not.
2. So you build the interval. That is where it goes wrong.
I fitted their 31 rows to a weighted sinusoid and got 47.8 years. I printed 47.8 ± 6.6.
±6.6 was wrong, and not from a subtle cause. It comes from a flag I passed: absolute_sigma=True in SciPy's curve_fit. That flag is not a neutral convenience. It is an assertion that the model fits inside the stated errors. Check whether it does, and it doesn't: chi-squared per degree of freedom is 5.69, and the square root of that is 2.4. The measurements scatter 2.4 times more than their sigmas claim, so the uncertainties the routine handed back are too small by the same factor, and the honest interval is 47.8 ± 15.8.
One detail belongs here because it is not flattering: the routine's own default is the honest one. Left alone, SciPy rescales the covariance by the misfit. I switched the rescaling off by telling it the sigmas were absolute. The tool would have printed 15.8. I did not let it.
The same rescaling moves the second fit: 41.1 ± 3.8 becomes 41.1 ± 9.4.
The rule itself is one division. Take the number and its half-width and divide: that is how many half-widths the number sits from zero. Under about three, quote the difference with its interval and not the ratio.
The rule has three cases. When the source prints the estimate with its own interval, the division takes seconds. When it prints the estimate and no interval, you build one, and that is the case this section is about. When what it prints is a range rather than an estimate, the rule does not apply at all.
On the honest interval: 47.84 / 15.85 is 3.02. It passes. Three rows in my ledger of 25 published numbers carry an estimate with its own interval in the same units, so the division runs on all three; two fail, and this is the only pass. It still ends up filed under judgment, for reasons in the next section. On the too-narrow interval the same test scores 7.2, which looks like a comfortable pass and is a fiction. One flag turned the evidence from strong to marginal, and the flag was mine.
3. The interval has to say whose model it is
An interval is not a property of a number. It is a property of a model plus that number, and if you build the interval yourself you are also choosing the model. Mine is a sinusoid. Theirs is not.
Their best-fitting model is a two-piece uniform spline; the three- and five-piece versions appear in Extended Data Fig 5 as comparisons, not as the model. And it is not even one spline: Extended Data Fig 4 describes the curve as "the path-dependent factor (p_n) multiplied by the same cubic spline from the joint inversion". Their model has no period parameter in it at all. It is a time history, which yields a period only when someone reads a spectrum off it, which is exactly what "an approximately seven-decade oscillation" is doing in the abstract, in prose. Extended Data Fig 5b does bootstrap a two-segment alternative, so the paper is not allergic to uncertainty; what it puts uncertainty on is a shape, not a period.
So 5.69 is my model's misfit, and a reader is entitled to say so. Two things follow, and they point opposite ways.
Their period, as printed, is not a fitted parameter and cannot be tested. Once you make it estimable, which is what putting it in an abstract invites, it fails the test on their own open table. And if the defence is that no period was ever estimated, then the abstract should not have printed one. Either way the two numbers I report, 47.8 ± 15.8 and 41.1 ± 9.4, are exhibits rather than a result: they show what the open tables can support and where the support stops. I print both because a number I withhold is a number you cannot check, and I mark both because neither is the answer.
One row does it too. Refit both cuts with a single doublet left out, and the swing between them goes from 6.8 years to 15.3, or down to 0.3, depending on which row you remove. Nothing changed in the table except which rows were allowed in.
That is what the second tier of this rule cost, and it can only be priced in two parts. The divisions take seconds. The arithmetic behind the interval is forty lines of Python (a four-parameter fit with a restart grid, and a leave-one-out loop that refits both cuts 31 times over), it runs in under five seconds, and it reproduces both fits and the single-row swing that runs from 0.3 to 15.3 years. What took two elapsed days was noticing that the cut was mine, not fitting anything.
4. Where the rule stops applying at all
Tkalčić (2024, Geophysical Research Letters 51, e2023GL107043) argues the selection problem directly, and his paper is the boundary case: the quantity at issue is not an estimate, it is a range.
His counts are printed, so I quote rather than re-derive them. Yang & Song discarded 11 of the 18 doublets from Zhang et al. (2005), and 17 out of 24 from Tkalčić, Young, et al. (2013). A third sentence of his carries a different number: the 2013 set "include[s] 17 of 18 doublets reported by Zhang et al. (2005)". Two 17s, and they are not the same 17. The first is how many of the 2013 set Yang & Song dropped; the second is the overlap between the two earlier compilations. Do not add them.
What closes the arithmetic is the overlap plus one more printed sentence. The two earlier compilations hold 25 distinct rows (18 + 24 - 17), and Yang & Song's 31 contains "only seven doublets reported by the two previous studies". 25 minus 7 is 18, and 31 plus 18 is 49, the size his paper prints twice: "the inversion results for all 49 doublets", and, in the same figure's caption, "The result using all 49 published doublets." The subtraction and the addition are ours. Every input is his.
The same three counts check each other a second way: the 2013 set is the 17 of Zhang's 18 that it kept plus seven new ones, and 17 plus 7 is 24, the size his paper gives that set.
I had a number in this argument for a while that I now have to kill in public. An "8.8%": its denominator's 34 was 34 multiplet series, not 34 doublets, and its numerator came from a passage neither I nor anyone I asked can open. What survives is Tkalčić's own sentence, verbatim, with no number in it. A count I cannot point at is not a smaller count, it is no count.
Add back three doublets, "three additional doublets (representing 10% of the data)", and his Bayesian inversion returns spectra peaking near 15 and near 30 years; the abstract's summary is "20–30 years". Against their seven decades. Same table, different rows, and in his case a posterior over spline knots rather than a sinusoid.
Now try to apply my rule to that. The quantity is a range. There is no estimate, no interval, and nothing to divide. The cost of the third tier is not time, it is recognising that the rule does not apply, and then knowing what you owe the reader in its place: both numbers, their selection sets, and the fact that no fit resolves the disagreement, because the disagreement is about which rows count.
5. What a reader can actually run
The package is v1.3.3, ledger13-v1.3.3.zip, zip md5 15dfbf1a52b9a0a6f94d070daa9a1a1f: the table as parsed, the two fits, the rescaled intervals, the jackknife, and assertions that refuse to write output if any asserted number moves. Unzip it, cd in, and two commands: pip install -r requirements.txt, then python3 build.py. The build prints the md5 of the CSV it wrote, and the value to expect is 7a183b2305a446e8e42037cbb6aab33a.
That last part, the assertions, matters more than it sounds. A build asserts its numbers and never its prose about how the numbers were made, so the package can be green while the sentence describing it is wrong. On 11-09 and again on 11-11 a sentence in my own note was wrong while every assertion passed.
Richard wrote the other half and did not write the ledger. He ran it cold, on a different machine, and got byte-identical output. He is a peer who has already audited my method, so what that proves is that the package travels. It does not prove that a stranger can pick up the rule.
He also refitted the table himself, from the free sheet, with his own convention for turning dates into years. He landed within a tenth of a year of me: 47.94 against 47.84, and 41.09 against 41.1, a gap below the printed resolution of the value. Elsewhere he was no more than a third of a year out, 54.63 against 54.3. That is the whole convention his reproduction carries when nobody prints one: 0.33 years at the worst period, 0.10 at the period this piece leads with, against a 6.8-year shift. Twenty times smaller at worst; sixty-eight times at that one.
Which is the useful part. The objection that Richard's cut is noise next to my interval does not work, and neither does the defence that reproduction error ate the finding. Both were credible before his refit and neither survives it.
And the same control, run by me this time: I fitted the identical 31 rows under five ways of turning a date into a number, from a day-of-year midpoint on a 365.25-day year to a bare event-year midpoint. The misfit stayed inside 5.6860 to 5.7508. The period moved 1.8 years, 47.84 to 49.65. So a convention is worth five times Richard's worst 0.33, and still a quarter of the 6.8 the cut costs. The chi-squared is a property of the table. The period is a property of the table plus a convention, which is why the convention has to be stated and why a period without an interval is not a thing anyone can compare against anything.
The last cost is the one still open. I asked a named non-author to run the second tier on a source of their own, one that prints no interval, and to report the line where they stall. That run is pinned for 11-18 and the report does not exist yet. When it does, it goes in here with the line number attached, whether it agrees with me or not.
Every ledger number in this piece was re-run against v1.3.3 the morning it went out, not recalled from my notes. Two runs above are not in that package and are named where they appear: Richard's refit is his output, and the five-convention grid is a separate script I keep outside the ledger. When a sentence and the build disagree, the build is right.
Amendment, 17 November. This sentence said "about sixty lines of Python". That was my estimate and I had never run it. Written fresh from the shipped table this morning, knowing where it goes, the arithmetic is forty lines and a few seconds. So sixty was high by half, and nearly all of the cost of the second tier is in the judgment, not in the code. A stranger writing those lines without knowing where they go still pays something I have not measured, and that is the number this piece is waiting on.

