What we shipped
A proactive next-24h contagion watchlist. The contagion-chain model shipped earlier this week is descriptive: it answers “given country A blocked today, P(B follows in 7d) = X.” This new endpoint flips the question around and runs it live: given the triggers that actually fired in the last 48h, which other countries are most at risk of blocking in the next 24-48h?
This is the difference between “here is a static contagion table” and a watchlist that updates every 6h.
How it works
-
Pull active triggers — any country with a confirmed
censorshipormixedincident in the last 48h. Dedup to one trigger per (country, day). - Score every candidate follower at horizon = 3 days (the shortest supported horizon — closest proxy for “next 24-48h”). Each follower gets one probability per active trigger.
-
Aggregate via noisy-OR: combined probability that the follower
blocks given any active trigger fires is
1 - ∏(1 - p_i). This caps the combined probability at 1.0 and assumes triggers act ~independently. -
Base-rate-adjust: subtract
base_rate_daily × 3so countries that block every day anyway don’t dominate. Final score ismax(0, noisy_or - base_prior_3d). - Self-exclude: countries that were themselves triggers in the last 48h are excluded from the watchlist.
Endpoints
-
GET /v1/atlas/contagion-watchlist— top-K candidates + active-trigger meta. Optional?top=N(max 50). -
GET /v1/atlas/contagion-watchlist/info— meta only (small payload for cards / dashboards). -
GET /v1/atlas/contagion-watchlist/{cc}— per-country attribution: which active triggers are pushing this country up the list and by how much.
Example call
curl https://api.voidly.ai/v1/atlas/contagion-watchlist | jq .
Honest caveats (baked into every response)
- Inherits the underlying contagion-chain model’s moderate AUC (~0.67). This is signal, not a deterministic forecast.
- 24-48h is a tight window. Most observed contagion patterns are 3-7d. The endpoint uses horizon = 3 days as the closest available proxy.
- A country on the watchlist does not mean it will block. The score is a relative ranking. Absolute calibration of the underlying pairwise classifier is moderate (Brier ~0.14 at 3d).
- Trigger attribution is probabilistic, not causal. Two countries spiking around the same time can reflect coincidence, shared upstream outages, or regional politics.
- Noisy-OR aggregation assumes triggers act independently, which overestimates risk when multiple triggers share an underlying driver (e.g. a regional election cycle).
- Base-rate adjustment is a heuristic, not a matched counterfactual.
- When zero triggers have fired in the last 48h, the watchlist is empty by design.
Why this matters
Most censorship-forecasting work is either single-country (“Iran 7-day shutdown risk”) or static pairwise tables (“OM → JO is a high-confidence chain”). Neither tells a journalist or operator which 5 countries to watch right now. This watchlist is the connecting layer: it consumes the trigger stream that’s already in the incident pipeline, runs the existing pairwise classifier, and surfaces a small ranked set with per-trigger attribution — all with the same honest-caveats stack the underlying model carries.