Munim Technologies

Research/Earth

A 160 ms change at the core-mantle boundary is not in the public data

The only reported decade-scale change of an ultra-low velocity zone rests on one station. Re-measured from the public records with a pre-registered test, its 160 ms SKS advance does not appear: the change is +28 ms, in the opposite direction.

Nov 21, 2026 · 7 min read

In 2025, Zhang and Wen reported in *Nature Communications* that an ultra-low velocity zone (ULVZ), a thin patch of partly molten rock on the core-mantle boundary 2,900 km down, had shrunk or moved by tens of kilometres in under ten years. It is the only published decade-scale change of such a structure, and it rests on one pair of Aleutian earthquakes (July 2000 and December 2009) recorded at one station, IU.SDV in Venezuela. The key observation is that the radial SKS wave arrived 160 ms earlier in 2009. We re-measured it from the same public records in a test whose rules were published before any waveform was downloaded. The advance is not there. The SKS-minus-S time changed by +28 ms, the opposite sign, and the two SKS waveforms are nearly identical.

-160 ms

reported change of radial SKS, 2009 vs 2000

+28 ms

our measurement (SKS minus transverse S, metadata rotation)

0.99

correlation of the 2000 and 2009 radial SKS waveforms

0.55°

measured sensor rotation between the events (metadata says 3°)

Why we looked

If the deepest mantle changes on human timescales, it would reshape how we think about heat leaving the core and the material that feeds mantle plumes. A claim like that needs every non-Earth explanation ruled out. Reading the station's metadata, we noticed that IU.SDV's horizontal sensors were recorded as re-oriented by 3 degrees on 2009-09-18, three months before the second earthquake, while the instrument response stayed the same. The change was seen only on the radial component, and a rotation error mixes part of the transverse motion into the radial. The paper rules out clock errors but does not discuss orientation. That made a cheap, sharp test: does the change survive the orientation question?

What we did

We downloaded both records from EarthScope, removed the instrument response, and filtered them to 0.4-2 Hz displacement as in the paper. We relocated the 2009 event relative to the 2000 event using mantle P waves at 13 global stations (it lies 0.6 km away; the location correction is 0.3 ms). Instead of absolute times, we measured the change of SKS relative to phases the paper reports as unchanged: the transverse S-Scd-ScS wave, and the vertical P wave. This cancels origin-time and clock errors. We did it three ways: rotating each record with its own metadata, rotating both with nominal north and east, and rotating the 2009 record by the relative rotation measured directly from long-period waves. The decision rule, written in advance, called the change reproduced if either common workflow gave an SKS advance of at least 100 ms. The plan is public.

What we found

All five quality gates passed: the 2009 event is a verified member of the same source cluster (12 stations with P-wave correlation of at least 0.95), SKS signal-to-noise is above 20 in both records, and the transverse S waves correlate at 0.99. Under every workflow and both references, the 2009 radial SKS moves by between -13 and +48 ms. The pre-registered grade is not-reproduced.

Change of the radial SKS arrival at IU.SDV, 2009 minus 2000

Negative means earlier in 2009. The paper reports -160 ms. Location-corrected cross-correlation lags, 0.4-2 Hz displacement. Source: results/sdv23-v2/sdv23_pairs.csv

The orientation question turned out not to matter. Matching the long-period horizontal waves of the two earthquakes shows the sensor turned by only 0.55 degrees between them, not the 3 degrees in the metadata, and switching between rotations moves the result by about 19 ms. There is simply no 160 ms shift to explain. One window does move early in 2009 under every reference: the start of the radial S-Scd-ScS wave (-32 to -80 ms), the same direction as the paper's reported -80 ms. We did not pre-register a grade for it, and the middle of that wave does not match the paper's -30 ms.

Every window, 2009 minus 2000, relative to transverse S (metadata rotation)

Lags from 3.5-s to 13-s cross-correlation windows set on the 2000 record. Source: results/sdv23-v2/sdv23_pairs.csv

We also looked for other earthquakes in the same cluster that could date a change. Six other events within 12 km were checked; only a January 1996 event is a member, and its SKS at SDV is too weak to measure. Our forecasts, made before the data, gave this outcome 35%. We had expected the change to survive.

Pre-registered forecasts and outcomes

ForecastProbabilityOutcome
Primary test is testable55%testable
Change survives / not reproduced50% / 35%not reproduced
Sensor physically rotated by about 3 degrees45%no: 0.55 degrees
Another cluster member verified40%yes (1996), too weak at SDV
From findings/2026-11-21-prereg23-results.md.

What it means

With the public records and standard processing, the waveform evidence for the only reported decadal ULVZ change is much weaker than reported. That fits a 2026 study by Cui, Wang, Vidale and colleagues, which found the core-mantle boundary mostly stable in 231 repeating earthquake pairs. It does not yet show the paper is wrong. The paper does not say how the 160 ms was measured, and an onset-based reading of a slightly broader 2009 pulse could differ from our whole-pulse correlation.

We tested that idea on 22 November, and the test could not be graded. Our pre-registered onset definition needs the wave's envelope to fall below 20% of its peak within 3 seconds before the peak. At this station and frequency band, radial SKS rides on a longer swing and never does, in either record, so we report the result as untestable rather than change the rule after seeing the data. The same run measured, without a grade, how much broader the 2009 pulse is: radial SKS by 4%, P and S at SDV by 0.5 to 2%, and P at 12 other stations by 0% (median). Under the plan's own arithmetic, a 4% broadening moves an onset by tens of milliseconds, not 160.

The first follow-up runs the identical measurement on the authors' own published traces. Its first run compared their unfiltered 2000 traces with their band-passed 2009 traces because of a file-pairing error in our code, so it could not be graded. We published that result as it stands and queued a corrected run. Two more tests are queued, both with rules published in advance. One measures SKS at other stations whose waves enter the core near SDV's entry point, where the paper places the ULVZ. The other measures SKS at SDV for other earthquake pairs whose waves leave the core near the same exit point, the paper's alternative location.

More research