Research/Earth
A hint of inner-core change at Yellowknife was the recording hardware
Two pre-registered tests show that a time trend in inner-core waves at the Yellowknife array, seen in repeating earthquakes whose sources were verified alike, comes from the array's 2013 hardware change and not from Earth's core.
Nov 19, 2026 · 6 min read

Earlier today our open repository reported a weak but positive trend: among repeating earthquakes near the South Sandwich Islands whose sources we had checked to be alike, the inner-core wave recorded at the Yellowknife (YKA) array in Canada differed more the further apart in time two events were. That is what a changing inner core predicts. It is also what a change of instruments predicts, and YKA replaced its recording system in 2013. We wrote down two tests and their decision rules before running them. Both say the same thing: the trend comes from the hardware. Inside one recording system there is no trend, and once the measured hardware effect is removed, the trend disappears.
This matters because repeating earthquakes recorded at YKA and at ILAR in Alaska are the main evidence in the 2024 *Nature* study by Wang, Vidale and colleagues that the inner core recently reversed its motion relative to the mantle ("backtracking"). Any test of inner-core change that compares YKA waveforms across 2013 needs a hardware correction. Here we measure that correction directly.
The question
Our earlier pre-registered test H13 scored every published YKA repeater pair by how alike the two earthquakes look in mantle P waves at eight stations that never see the core. Among the 328 pairs (42 families) whose sources were verified alike, the difference D in the 1-6 s coda of the inner-core phase PKIKP grew with time apart: Spearman rho = 0.158, family bootstrap 5th-95th percentile 0.061 to 0.255. YKA's pair table labels each record by the system that recorded it: the legacy system (20 samples per second, until 2014-04-02) or the current one (40 samples per second, from 2013-06-26). Pairs far apart in time almost always mix the two.
What we did
Pre-registration 17 repeated H13 on pairs recorded by the same system only. Pre-registration 18 used a gift in the station records: from June 2013 to April 2014 both systems recorded side by side on the same sensor type at the same sites. For 37 South Sandwich Islands earthquakes in that window that passed the array's quality gates, we built a legacy beam and a current beam of the same wave and measured their coda difference D_hw. That is the hardware effect alone, with the earthquake and the Earth held fixed. We then removed it from every cross-system pair and re-ran H13. Both scripts first had to reproduce H13's numbers exactly; both did.
37
earthquakes recorded by both YKA systems and passing the gates
0.0117
median hardware-only coda difference D_hw
-0.019
H13 time correlation after removing it (was 0.158)
15%
our pre-registered forecast for this outcome
What we found
Inside one generation, the time signal is gone. Among 179 same-system verified pairs from 31 families, rho = 0.028 (5th-95th percentile -0.090 to 0.107), which meets the pre-registered rule for no time signal (rho below 0.10 and 95th percentile below 0.20). The current system alone, 156 pairs from 2015 to 2026, shows rho = 0.067. Across generations the correlation is -0.054. On the pooled set the trend appears only because pairs ten or more years apart are almost all cross-system: 117 of 121.
Time correlation of the inner-core coda difference, by recording system (pre-registration 17)
Spearman rho
The hardware effect is small for any single pair. The two systems' beams of the same earthquake overlap almost wiggle for wiggle (median D_hw 0.0117, range 0.004 to 0.021), against a median pair D of about 0.02 to 0.07. No event needed a polarity flip, and the two systems' timing differs by a median 0.087 s, which our alignment removes. But H13's trend was itself a small shift in ranks. Removing a term this size from cross-system pairs takes rho from 0.158 to -0.019, and to -0.035 with the largest plausible term. The pre-registered verdict is hardware-explains.
H13 time correlation before and after removing the measured hardware term (pre-registration 18)
Spearman rho
Median coda difference by time apart
| Years apart | Same-system pairs | Median D, same | Cross-system pairs | Median D, cross |
|---|---|---|---|---|
| 0-3 | 86 | 0.023 | 0 | - |
| 3-6 | 56 | 0.023 | 9 | 0.027 |
| 6-10 | 33 | 0.035 | 23 | 0.066 |
| 10-15 | 4 | 0.034 | 63 | 0.036 |
| 15-30 | 0 | - | 54 | 0.032 |
Not everything fits neatly. In the 6-10 year bin, cross-system pairs differ by 0.030 more than same-system pairs, about 2.6 times the hardware term. And the late coda (3-10 s) keeps a weak correlation with time after correction (rho 0.112, 5th percentile -0.008), which we graded inconclusive as a secondary window. Neither changes the primary grade, and inside one system the late coda shows no trend either (rho 0.037).
What it means
We forecast, before the data, that the trend would survive the hardware test (50%) and gave the hardware explanation only 15%. We were wrong. H13 is not evidence for temporal change of the inner-core path to YKA. What still stands is the single-family result we published on 18 November: the M03-24 family's core waves split its eight events into two groups while mantle waves do not. That split puts the 2003 event, recorded by the legacy system, with the 2020-2025 events recorded by the current one, so it is not a hardware split. Our next steps are a matching hardware measurement for ILAR's 2018 digitiser change and a check of whether larger earthquakes make near-antipodal stations usable.


