On 2026-04-17 at 02:00 UTC, the Voidly Sentinel forecast for Lebanon crossed its alert threshold for the first time in this run: the calibrated next-7-day shutdown probability sat at 0.075, against a threshold of 0.05. Sentinel runs every morning at 02:00 UTC on the production Vultr server. From that morning forward, Lebanon stayed above threshold every single day for the next 28 days — even climbing as high as 0.118 by May 5.
On 2026-04-24 at 12:25 UTC — six and a half days after the second forecast row in the run (2026-04-18) — IODA flagged a critical BGP-level connectivity drop on the North Lebanon regional entity. The drop persisted long enough to publish as incident LB-2026-0098. Voidly Atlas captured the IODA evidence at 18:00 UTC the same day and minted the permanent incident ID. The forecast had been elevated for the whole week prior.
The pattern repeated. The model stayed above threshold; IODA fired again on 2026-04-25 (incident LB-2026-0100), 2026-04-30 (LB-2026-0105), and 2026-05-03 (LB-2026-0108). The forecast was elevated for every one of those events. The model started to come back down to baseline around 2026-05-16 once the recent-shutdown signal aged out of the 30-day window.
What this case does — and does not — show
What it does show is the alert mechanism working end-to-end: a threshold cross on a specific morning, sustained elevation, a real citable IODA event inside the window, and a permanent incident ID anyone can verify. That is the pipeline — forecast → threshold cross → evidence — working as designed.
What it does not show is predictive skill. Lebanon’s internet has been chronically unstable since 2019 and IODA reports outages there several times a month, so a forecast that sits just above a low (0.05) threshold will — by base rate alone — often be “above threshold” when an outage happens to land. The shutdown-risk model’s own validation is blunt about this: its within-country timing (the “when”, as opposed to the cross-country “which country”) is near chance. So treat this as a worked example of how an alert flows to citable evidence — not as a claim that Voidly can time Lebanon’s outages. We would be overclaiming if we did, and the model’s published metrics would catch us.
The honest caveat
Sentinel is not always right. On the same 30-day window that produced this case, the model’s per-country precision in Lebanon was 0.54 (15 true positives / 13 false positives) at recall 1.00. In other words: when the model was elevated for Lebanon, an outage followed roughly half the time — barely better than a coin flip for a country that is near-constantly elevated. Across all 30 watched countries, the rolling-window precision is 0.70 at recall 0.36. We publish the full reliability data at /sentinel/backtest and the live calibration curve at /sentinel/calibration. This case study is one event, not a guarantee — and a single favorable anecdote is exactly the kind of evidence our own honest evals tell you to distrust.