We retract our earthquake precursor.
This page used to claim that ocean microseism amplitude rises before large earthquakes, with a lead time of about ten hours. Our own verification found that result to be an artifact. The signal was not the ocean. It was the earthquakes recording themselves. We are leaving this page up, at its original address, with the correction attached.
What we claimed
That across thousands of historical M≥5.5 earthquakes, a secondary-microseism amplitude metric rose systematically in the hours before rupture, and that the effect was large, global, and scaled with magnitude. We proposed pre-rupture dilatancy as the mechanism.
What was actually wrong
The quantity we were measuring was not a band-limited ocean microseism amplitude. It was a whole-day, broadband noise-exceedance metric: a statistic that rises for any local seismic energy, including the earthquake it was supposed to precede. On a daily-resolution feature, a “ten hours before” window lands on the same calendar day as the event most of the time. So the measurement was substantially concurrent with the rupture rather than ahead of it.
Three things confirmed it. The apparent signal peaks on the event day itself and decays afterwards along an aftershock tail. A “precursor” that continues after the event is not a precursor. After removing aftershock sequences, the strictly-before-the-event residual collapses and stops scaling with magnitude, which is precisely the property we had cited as evidence the effect was physical. And when we re-ran the test on a genuinely band-limited hourly metric computed from raw waveforms, the pre-event signal was decisively excluded, while the same pipeline recovered the earthquake itself at very high significance: the positive control that makes the null trustworthy.
What survives
A small residual before the event, consistent with ordinary foreshock and sequence seismicity. That is a real and well-documented phenomenon, and it is not what we claimed. It is not an ocean-microseism precursor, and it does not support an earthquake alarm.
What we did about it
The pre-registrations covering this result were annotated as disconfirmed, never rewritten: their Bitcoin timestamps still verify against the bytes that were sealed. The paper draft was withdrawn rather than submitted. The operational predictor no longer writes forecasts to our prediction ledger. And this page was rewritten rather than deleted, because removing it would have left the original claim circulating with nothing attached to correct it.
Why we are publishing this
A forecasting record is only worth something if the failures are in it. We pre-register our hypotheses, anchor them so the date cannot be moved, and publish the result either way. This one failed. Finding it took an adversarial re-test of our own strongest result, and the honest outcome of that work is a retraction, which is the part of the method that makes everything else we publish checkable.
DeepMap AI · retraction issued 24 August 2026 · this page updated 9 September 2026.