Live record
FISHER — Fast Inference of Seismic moment with Heteroscedastic Error Reporting. One station, one vertical component, one window of P-wave motion, and a magnitude with an uncertainty attached to it. The uncertainty is the reason for the mouthful: it is not a constant, it widens with magnitude and with how hard the record is, and it is calibrated rather than assumed.
Every event above Mw 5.5 with adequate signal is scored automatically, within minutes of the origin, and published here. The Global CMT solution — the reference this was trained to reproduce — is appended when it arrives, and the record then scores itself.
The record opens with a backfill. Everything from 1 January 2026 to the day the pipeline went live was scored after the fact, against earthquakes whose reference solutions already existed. Those events are marked backfilled and counted separately, because a prediction made after the answer is published is not a prediction. They are here to establish a baseline against a few hundred events rather than a handful. Everything from mid-August 2026 onward is scored before its reference exists, and carries no such mark.
Global CMT publishes a quick solution within hours for most events, and revises it later. That is the window this has to beat: an estimate is only prospective if it was made first, and the timestamps on every card say when.
Three window lengths are reported: 60, 90 and 120 seconds after the P arrival. Mwp, the classical method these are measured against, integrates 120 seconds — so the 60-second column is half the wait for an answer that is still ahead of it. Shorter windows exist in the sensitivity analysis but are not operational products and are not shown.
Stations that could not contribute are listed with the reason. Most are simple geometry: outside the 20–100° range there is no direct P to measure. The rest failed a signal-to-noise floor, and publishing those is deliberate — a record that quietly dropped them would over-represent good recording conditions and flatter itself.
Windows
Baselines
Sort
Loading the event record…
Shaded bands are one standard deviation from the model's own predictive distribution. Mwp and mB_BB carry no uncertainty estimate, which is why they appear as bare marks. The axis is linear in Mw, which is already logarithmic in seismic moment: one whole unit is a factor of about 32 in M₀. It is scaled per event so the bands stay legible, but the span snaps to a fixed set and is printed beside each axis, so widths stay comparable between events.