Many internet shutdowns are reactive: a protest builds, the story breaks in international media, and within hours connectivity drops. The pattern is familiar enough that Access Now's Keep It On reports cite it directly — but until today Voidly Atlas's forecasting stack didn't read protest-mention data directly. The pre-protest correlator changes that. We pull daily GDELT counts of PROTEST+RIOT news mentions for every censorship-heavy country we track, compute a 30-day rolling z-score per country, and then test the hypothesis that "today's protest z-score predicts a shutdown within 48 hours."

Across 16 countries with enough overlap between GDELT data and confirmed shutdown labels in the 2026-01-20..present window, we found 2 countries with statistically significant correlation at our promote-floor (Pearson r ≥ 0.30, p < 0.05): Myanmar (r = +0.41, AUC = 0.85, n = 5 shutdowns) and Nigeria (r = +0.34, AUC = 0.97, n = 2 shutdowns). Saudi Arabia sits right at the threshold (r = +0.30) but its AUC is near-chance (0.48) — the correlation comes from a tail of high-z days rather than a clean separation, so we don't count it. Two more countries with mild positive signal (Russia r = +0.10, Egypt r = +0.10) confirm the direction-of-effect but aren't individually significant on this sample.

Myanmar is the clean example. From 54 days of overlap with 5 confirmed censorship events, the protest z-score on the day-of and one day before each shutdown was on average 1.8 standard deviations above the country's own 30-day baseline. AUC = 0.85 means that if you took a random "shutdown next 48h" day and a random "no shutdown next 48h" day, the z-score would correctly rank them 85% of the time. Nigeria's correlation rests on only 2 shutdowns — statistically significant by p-value but operationally fragile, and we say so explicitly in the endpoint response.

Because we only cleared 2/5 of our promote-floor countries, the model ships as "promoted": false. We expose all 16 countries' results plus current protest z-scores anyway, because the underlying signal — protest mentions today, shutdown tomorrow — is a real pattern even when it doesn't generalize across every country we test. As GDELT history grows past the 2026-01-20 start, more countries will accumulate enough shutdowns to clear the bar. The next bump in significance count will likely be Russia and Egypt, both of which trend positive but lack the sample size today.

Honest caveats we don't hide: GDELT counts news mentions, not protest events. A spike could be coverage of someone else's protest, or a single viral story amplified across the English-language press. The per-country 30-day baseline is narrow enough that sparse-news countries (e.g. Turkmenistan) get inflated z-scores from low variance. Confirmed-censorship labels exclude IODA-only disruptions, which means weather-driven and BGP-driven outages don't pollute the training signal — but they also shrink the positive class. And correlation isn't causation: a protest spike and a shutdown might share an upstream cause (election week, policy announcement) rather than one triggering the other.

What this is, what this isn't: it's a supporting signal — a 24-48h early-warning input that can feed into the broader forecast stack. It is NOT a standalone predictor. For Myanmar specifically we're confident enough to surface "elevated GDELT signal" as context next to a current_risk query. For the other 14 countries the endpoint returns the numbers honestly with the significance flag set to false, so downstream consumers (journalists, the agent, future ML) can decide what to do with it.