Prediction Track Record

These are the subset of predictions produced by classifiers trained on real event outcomes (not synthetic distributions) plus physics-constrained reactive products. Broken out here because they represent the honest upper bound of current forecasting skill.

Fires only when ≥3 independent physics channels (seismic, electromagnetic, gravity, atmospheric, particle, oceanic, thermal, fluid) each cross a 2σ threshold in the same 15-min tensor snapshot. By construction, the expected false-positive rate at this gate is <3% for uncorrelated channels. This is the highest-precision, lowest-volume product in the ledger.
The gate is running on AX41 (13 * * * * cron); it has not yet produced its first ledger fire. That is expected behaviour at design-threshold: fires are supposed to be rare.

Tests whether the Cascadia subduction zone AND the southern San Andreas (UCERF3) are both elevated above their typical 7-day rate in the same window. Goldfinger et al. 2025 (Geosphere 21(6):1132) identifies 3 paired Cascadia+SAF events in 1,500 years from sediment-core turbidite analysis -- the 1700 paired M~9 + M~7.5 is the most recent example. The paper deliberately stops at correlation; this card follows the same discipline.

Prior derived from 3 paired co-occurrences in 1,500 yr turbidite stratigraphy (Goldfinger 2025). DB-backed

0.01/7d hard cap encoded — we cannot claim recurrence stronger than paleoseismic evidence supports.

Per-region tiered-horizon cascade for industrial shutdown protocols (smelters, chip fabs, rail). Each of the ~54 monitored earthquake regions independently emits up to 3 linked predictions per run — T-7d early warning, T-36h convergence tightening, T-24h industrial alert. Confidence thresholds are 0.60 / 0.75 / 0.85. Every tier carries the M0.5 band (or M0.25 once a region graduates on residual sigma ≤ 0.30 with ≥ 30 matched obs), a 95% CI, and a SHA-256 hash-chain entry.

Last 30 days, post-PR#28 cadence-artifact fix (Apr 20). Bulk Old Faithful is now logged as one prediction per predicted eruption (was every ~10 min; the old number was a cadence artifact, not skill). Specific geysers (Daisy, Beehive, Riverside) are scored per-eruption. Small N (n<10) and irregular-cadence research-tier geysers are badged so 100% on n=6 doesn't read the same as 100% on n=60.

Trained on the unified Earth State Tensor (9,687 cross-physics features, hourly snapshots). Inputs include real-time gravity, solar wind, electromagnetic, oceanic, particle, fluid, atmospheric, and seismic measurements. Outputs are calibrated probabilities of M5.5+ events, cross-validated leave-one-out for honest skill reporting.

These are skill metrics measured against the new Apr 15 bounds-first matcher. Full-credit means the observed event's magnitude met or exceeded the prediction's expected range. Nowcast AUC measures whether our risk score rankings correlate with observed activity.

Each confirmed prediction is graded A-F based on distance accuracy (60%) and lead time (40%). All confirmations are matched against authoritative sources (USGS, EMSC/GFZ, IRIS, GeyserTimes).

Data provenance: all observations verified against USGS Earthquake Hazards Program, European-Mediterranean Seismological Centre (EMSC/GFZ), NOAA SWPC, and GeyserTimes citizen science data. SHA-256 hash chain ensures predictions are tamper-proof and were logged before the confirming event occurred.