Most risk vendors assert their accuracy. We publish ours, with timestamps, and leave it open to inspection. Below are early-warning calls verified against official sources — and the live detection log the engine writes as it runs. No precision figure to take on trust. Check the dates against when the headlines actually broke.
Scorecard · live from the journal
The confusion matrix, in the open
One episode is one situation in one country: the journal holds a row per country-day, and days within two weeks of each other are folded into a single call, so a six-week situation counts once, not six times. Closed-loop counts straight from the server-authored prediction journal — including the calls that did not confirm and the escalations we stayed silent on. The honest denominator, not a cherry-picked numerator.
—
Confirmed episodes
+12.9%
7-day skill vs the country's own base rate (95% CI +4.5…+20.4, n=146)
—
Episodes in flight
—
Episodes that did not confirm
—
Misses tracked
—
Precision (scored)
Loading live scorecard from the journal…
Confirmed early calls
What we withdrew, and why
This section used to show lead times — "flagged 20 days before mainstream coverage". On 6 October 2026 we tested those numbers against a baseline for the first time and withdrew them. The lead times stopped dead at 31 days with nine of 24 piled against that ceiling: the shape of a capped query, not of foresight. And the fourteen countries the calls sit on contain an armed-conflict entry in an independent chronicle in 84% of random 31-day windows, against our 48% precision — naming those countries on any random date would have scored higher.
Nothing replaces them until a null model is attached and published beside the figure. The forecast numbers below are measured against the country's own history and carry their confidence interval.
Every confirmed episode · live from the journal
The full closed record
Each row is one episode — a situation in one country, however many days it ran — that the engine wrote to the journal and reality later confirmed. Forward predictions still in flight are not shown here — those are the working product. Sensitive jurisdictions are reflected in the scorecard counts but not itemised (we report neutral escalation signals only).
Country
Flagged on
Lead time
Signal
Live detection log
What the engine is tracking right now
Detection lead-time — how long and how broadly the engine has tracked each signal. This is raw monitoring data, not a claim of confirmed outcome. Pulled live from the public feed.
Signal
Location
Tracking since
Days
Sources
Level
Loading live detection log…
How to read this
Dated at issue time. Calls are written to the record when made — not backfilled. The value is the gap between our flag and the headlines.
Verified against official sources. Confirmation comes from ReliefWeb, UN OCHA, WHO and equivalents — not our own say-so.
Small early sample — and we track our misses too. This is an early record still calibrating. Lead time on past calls is not a guarantee of future ones.
Lead time, not a verdict. An early signal buys your team time to look — it does not replace your own judgement or on-the-ground sources.
Aggregator, not media. Vigilo aggregates and attributes open sources; it does not write editorial and is not a news outlet.