Every case below is reconstructed from Vigilo's append-only prediction journal — timestamped before the outcome, scored after it. No retrofitting: the same journal also records every miss, and the totals are public.
Sudan produces little international coverage between flare-ups, which is exactly when conventional monitoring goes blind. Vigilo's covert-elevated tier exists for this case: it weighs the country's structural base-rate against unusually quiet signals.
Verification: public scorecard, confirmed_calls feed.
A single lucky call proves nothing. Myanmar shows the system repeating: three separate watch-tier calls issued 2026-06-05, 07-07 and 07-08 — each confirmed with a 30–31 day lead. Repeatability on the same country is what separates a model from noise.
Verification: public scorecard — Myanmar rows.
Most advisories say "typhoon season — reconsider travel" for two whole months. Vigilo's seasonal layer marks the storm corridor for central Vietnam and issues a usable answer instead: the safe window. Travellers got "aim for Oct 22–31, south is calmer than center" while agencies were still issuing blanket warnings — and each incoming typhoon is then tracked 2–5 days out on official forecast tracks.
Layers: risk_calendar (seasons) + GDACS multi-country tracks.
Not everything needs prediction — some risk is simply scheduled. The Duma election window (Sep 19–22, 2026) sat in Vigilo's public risk calendar months in advance, with practical guidance attached: complete in-country business and flights before Sep 15. The known future is the cheapest early warning there is — if your system bothers to read it.
Layer: risk_calendar.json — elections + curated geopolitical dates.
Recall against external ground truth is 58% (USGS M6+ benchmark) — we catch most, not all. Precision 73.7% means roughly 1 call in 4 doesn't confirm. The journal is append-only: misses stay on the record, and the model re-validates itself blind every 8 hours (public/model_validation.json).
We publish these numbers because nobody else in this market does — and because a client who knows our error rate can actually plan with us.
30 minutes: your locations, your routes, our layers — on screen.
Book a walkthrough Inspect the raw scorecard