Research/Earth
Inner-core waves changed while the earthquakes themselves did not
In a pre-registered test, eight repeating South Sandwich Islands earthquakes from 2003 to 2025 sent near-identical mantle waves to seven independent stations, while their waves through Earth's inner core changed and changed back. The change is on the path through the core, not at the source.
Nov 18, 2026 · 6 min read

Seismic waves that cross Earth's inner core are the main evidence that the core rotates relative to the mantle, and in 2024 a Nature study argued that it has recently turned back. The evidence comes from repeating earthquakes: if the same source fires twice, any change in its waves must come from the path. That logic fails if the 'repeats' are not quite the same source. We tested exactly that for the best-studied repeating source near the South Sandwich Islands. In a test written down before the data were fetched, the earthquakes themselves did not change: their mantle waves stay alike at seven independent stations. Their waves through the inner core changed after 2003 and changed back by 2020. The change lies on the path through the core.
0 of 7
unseen mantle stations where the core-path split appears
8.0x
between-group vs within-group difference through the core at Yellowknife
1.19x
the same ratio in mantle P, median of 7 stations
2003-2025
span of the eight earthquakes, including one from December 2025
The question
The multiplet called M03-24 (59.03°S, 25.68°W) is the one that Zhang and Wen (2026, JGR Solid Earth) used to argue for changes of the inner-core surface rather than rotation. In our earlier, openly published work its eight earthquakes split into two groups by the late coda (3-10 s) of their waves through the inner core, and the split is the same at two arrays, ILAR in Alaska and YKA in Canada: {2003, 2020, 2022, 2024, 2025} against {2008, 2011, 2014}. That is the change-and-change-back that the backtracking timeline of Wang, Vidale and colleagues (2024, Nature) predicts.
The same pattern would appear if the 2008-2014 events had sat a little apart from the others, a separate source sub-cluster. Mantle P waves never enter the core, so they separate the two cases. If the split shows in mantle P, it is a source effect (S). If it does not, it arises on the path through the core (N).
The two groups seen through the inner core
| Group | Earthquakes |
|---|---|
| X | 2003-03-21, 2020-07-11, 2022-03-07, 2024-03-21, 2025-12-13 |
| Y | 2008-02-17, 2011-02-20, 2014-12-09 |
What we did
We wrote the rule down first (pre-registration 11). The code asked the EarthScope station service for every IU, II and GT broadband station 30-85° from the source and found 42. At each station it aligned all eight earthquakes on the observed P onset, kept those with clear signal and measured how different each pair's mantle P is 3-10 s after the onset. It then ranked the hypothesised X/Y split against every other way of splitting the same events. A station counted as a hit if the split ranked in its top 7%. With n usable stations, at most max(1, 0.2n) hits meant N and at least half meant S. Stations whose records we had looked at before were kept out of the decision.
What we found
Seven unseen stations in South America, Africa and Antarctica passed the quality gates. None of them showed the split: 0 hits, decision N. Our forecast before the run was N at 60%. The best station had p = 0.14; the rule needed 0.07 or less. Adding the South Pole station that had been looked at before gives 0 of 8.
How much more pairs across the groups differ than pairs within a group
Mantle P: how well the X/Y split is separated at each unseen station (p, lower = better separated)
We inspected the waveforms of every usable station before reading the decision. At all of them the stacked onset lies within 0.05 s of the reference onset, not at the edge of the search. At the quiet stations the eight records are nearly identical wiggle for wiggle out to 10 s. The late window holds the depth phases pP and sP, which react to small changes in depth or location. So a source difference large enough to cause the eightfold contrast through the core would very likely have shown up in mantle P too.
What it means
The same repeating earthquake sent nearly identical mantle waves in 2003, 2008, 2011, 2014, 2020, 2022, 2024 and 2025. Over the same years, its waves through the inner core to Alaska and Canada changed and changed back, and the 2024 and 2025 events stay on the 'back' side. This removes the main alternative explanation, a shifting source, for the clearest family in this debate. It also covers the December 2025 event, which no published study of this multiplet has used.
It does not say whether the inner core turned back or its surface changed and recovered; both predict a change on the core path. Telling them apart needs many families on one timeline. We have pre-registered that test next (pre-registration 13). It applies the same mantle-P source check to every published pair at both arrays. Then it asks whether the inner-core change among source-verified pairs follows the backtracking clock, a steady clock or neither.


