1. Ingestion

Scouts poll official feeds — USGS, NOAA SWPC, NOAA/NWS, NASA EONET, GDELT and FRED — and store the untouched provider payload before anything is interpreted. Every signal keeps its source URL and a revision number, so a corrected upstream record produces a new row rather than a silent edit.

2. Severity

Each domain maps its native intensity to a 0–100 scale in TypeScript: earthquake magnitude adjusted for depth, planetary K-index, GOES flare class, CAP severity/urgency/certainty, and trailing z-scores for economic stress. The same input always gives the same score.

Geopolitical stress is measured against each country's own history rather than an absolute threshold: coverage volume and tone from GDELT's conflict events are z-scored against that country's trailing 14 days of 15-minute windows, then blended with CAMEO Goldstein conflict intensity. A country with heavy routine conflict coverage therefore no longer scores highly on an ordinary day, and one that rarely appears is no longer invisible when it erupts. A window in which a country produced no conflict events counts as zero coverage rather than as no observation, so the baseline is what a country's normal quarter-hour looks like and not what it looks like on the days it makes the news. Tone is the exception: a window with no articles has no tone, so tone is averaged only over the windows a country appeared in, and a country too rarely covered for that is scored on coverage volume and conflict intensity alone. Below 48 windows of history there is no baseline at all and the reading falls back to absolute tone and volume. Every signal states which of the three produced its score, alongside the baseline mean, standard deviation and sample size, and the parties CAMEO named in the events.

3. Convergence

Signals within 750 km and 24 hours of one another are clustered; recency is weighted with a 18-hour half-life. Globally-scoped domains (space weather, economic stress) fold into every cluster. A cluster spanning more independent domains is multiplied up — 1 domains ×1, 2 domains ×1.4, 3 domains ×2.1, 4 domains ×3.2 — because independent corroboration is the entire signal.

4. Probability

Each domain and horizon has an empirical base rate. Cluster strength lifts that base rate toward, but never to, certainty; probabilities are capped at 93% and floored at 2%. Longer horizons are only opened for stronger clusters. A language model may phrase a statement but never computes the number.

5. Resolution

Outcome events are ingested independently of predictions. A prediction is graded on geography (30%), timing (25%), domain (25%) and intensity (20%). A combined score of 0.75 or above is a match, 0.5 to 0.75 is partial, and a window that closes below that is a miss. Misses stay in the ledger.

Intensity thresholds rise with the horizon, and the month and year horizons require qualifying events on two and three separate days respectively. A condition that a quiet year almost always satisfies — one moderate quake somewhere in twelve months — would resolve within days and flatter the scorecard without ever risking a miss.

6. Calibration

The scorecard reports Brier score, a reliability curve, hit rate, high-conviction precision and mean lead time. Any sample below 30 resolved predictions is labelled provisional. Early numbers will look unimpressive; publishing them anyway is the point.

7. The Daily Brief record

The public brief record begins on 2026-08-18. From that date, one brief per UTC day is published unedited: corrections appear in the next day's brief, never retroactively. A brief can only ever be withdrawn in full, with the reason and timestamp recorded and the original text retained internally — it is never rewritten in place. The launch-day brief of 2026-08-14 was auto-published under the superseded pre-0.2.0 scoring rules before this guarantee began, was sent to no subscriber, and has been withdrawn on that basis.

8. Known limits

Earthquake mainshock timing is not forecastable and Tremor does not attempt it; seismic outlooks concern aftershock and regional activity rates. Geopolitical tone data reflects media coverage, not ground truth. Economic stress indicators are lagged. Convergence is correlation, and correlation across independent physical systems is often coincidence — the ledger is what separates the two.