Our numbers, explained
Different questions need different measurements, so several rates appear across this site. Each one is defined here in one line, with the same definition used everywhere it appears. All of them come from the same public, Bitcoin-anchored ledger — misses included.
Area Skill Score (ASS) — 0.5 is no skill — the operator number: given how much ground we hold under alarm, do we catch more earthquakes than alarming that same fraction at random would? Measured Aug 2026: we hold about 85.6% of the world's climatology-weighted seismicity under alarm and catch 83.2% of events; random alarming at that fraction catches 85.6%. ASS 0.4877 — slightly below chance. This is the number that cannot be gamed by widening: alarm everything and it converges to 0.5 by construction. We publish it because anyone can compute it from this ledger, and a scoreboard that made them do that work would deserve the suspicion.
Brier Skill Score (BSS) — 0 is no skill — the pricing number: are our stated probabilities better than quoting the long-run base rate? Every figure is out-of-sample, from an expanding-window walk-forward. After per-model recalibration, most models are positive on an n-weighted mean of per-model out-of-sample BSS (Aug 2026; the live figure and per-model breakdown are on /track-record). Recalibration is monotone, so it improves BSS and leaves ASS unchanged — the two answer different questions, and only reporting both is honest.
8/8 M6.0+ (30 days) — coverage: of every M6.0-or-larger earthquake on Earth in the last 30 days (USGS catalog), how many were preceded by a logged forecast that later matched. Coverage says how much of the world's significant activity we had a call on — it is not a skill claim, because wide standing coverage of active belts makes matching easier.
60.4% when-issued (30 days, n≈600, as of Aug 2026) — conditional rate: of the forecasts we issued with a magnitude floor of M4.0+, how many verified with an event at least that large in the named place and window. The honest comparison: a naive model built only from historical catalog rates scores about 28.7% on the same predictions. This pair — our rate vs the naive rate — is worth scrutinizing, but it conditions on a forecast having been issued and says nothing about how much ground was under alarm to get it. That is what the Area Skill Score above measures, and it is the stricter test.
89% / 80% / 69% (14-day, settled 30-day, all-time; as of Aug 2026) — earthquake warnings that came true, on fully settled windows: every forecast counted, none removed, numbers final once the window closes. Different time ranges answer \"how is it going lately\" vs \"how has it always gone.\"
Full-magnitude rate (companion) — the strict version of the seismic hit rate: a forecast counts only if a matched event reached the full predicted magnitude floor. Standard accounting also credits events up to 0.5 magnitude below the floor as partial matches (about 1 in 5 recent matched earthquake forecasts were credited that way, 90-day audit, Aug 2026); both numbers come from the same live API, published side by side — neither replaces the other.
Contamination-excluded rate (companion) — the seismic hit rate recomputed with every forecast whose inputs were later proven fabricated removed from both the numerator and the denominator. Those forecasts are annotated, never deleted: the ledger and its hash chain stay byte-identical, and both numbers are published side by side.
Lead time (e.g. \"10.5 days ahead\") — the gap between when the forecast was logged (provable via its Bitcoin-anchored hash) and when the earthquake happened.