voidly atlas · live model

Shutdown Risk

Per-country 7-day internet shutdown risk forecast. Trained and validated on real, journalist-verified shutdowns from the Access Now #KeepItOn coalition — not on a proxy this project invented.

Model shutdown_risk_v9 · refreshed 2026-07-08 · 53 countries scored · 2 elevated · 7 on watch

Honest validation

Every probability below was scored against real KeepItOn shutdowns. The system reports its own performance with every response.

0.90

Cross-country AUC — predicting WHICH countries are at risk (Iran vs. Switzerland).

0.72

Within-country median AUC — predicting WHICH DAYS within a country (the real timing test).

13/23

Countries with usable timing skill (within-country AUC ≥ 0.65).

Score formula: isotonic(sigmoid(0.6*logit(shutdown_risk_v4_prob) + 0.40*logit(honest_forecast_v1_prob))). Threshold for "elevated": 0.100.

Data sources

Every number on this page traces back to a public, independent, citable dataset. None of the labels are invented by this project.

Ground-truth labels

Access Now #KeepItOn STOP

Journalist- and civil-society-verified internet shutdowns worldwide. 287 verified onsets in our prediction window (2022–2026), spanning 44 countries. The model trains and validates on these — not on a proxy this project invented.

KeepItOn data dashboard →

Measurement signals

OONI aggregation API

Measurements taken by volunteer OONI Probe users inside the countries being measured. 85,543 country-days across 60 countries, 2022–2026. Each day has real measurement_count denominators — not a synthetic proxy.

OONI public data →

Pipeline: scripts/pull-ooni-aggregation.py + KeepItOn STOP CSV are pulled daily, the ensemble retrains, and the leaderboard above refreshes. Source code: /scripts in the repo.

Today’s risk ranking

Top 30 of 53 countries sorted by 7-day shutdown probability.

What this is — and isn't. This reliably ranks which countries are heading into danger, and that skill is validated forward in time: in a strict past→future holdout the cross-country ranking holds at AUC 0.87 (country-ranking 0.90, 7.8× better than chance). It does not forecast which day a shutdown hits inside a given country. The within-country figure (median 0.74) is cross-sectional — measured with the model having seen the whole period; forward in time it drops to ~0.36 (below chance), because shutdowns cluster in time. Treat this as a forward-validated country risk ranking, not a within-country calendar oracle. Full numbers in the live model card.

#	Country	Probability	Band	Country risk	OONI signal	As of
1	MM	54.2%	elevated	41.6%	84.8%	2026-07-07
2	PK	27.6%	elevated	17.5%	81.6%	2026-07-07
3	RU	9.1%	watch	17.5%	48.1%	2026-07-07
4	TZ	9.1%	watch	6.7%	82.6%	2026-07-07
5	IN	6.0%	watch	41.6%	9.2%	2026-07-07
6	IQ	6.0%	watch	5.5%	70.1%	2026-07-07
7	IR	6.0%	watch	5.5%	59.2%	2026-07-07
8	JO	6.0%	watch	5.1%	52.5%	2026-07-07
9	SD	6.0%	watch	1.8%	82.6%	2026-06-19
10	OM	5.9%	low	1.0%	85.0%	2026-07-04
11	SY	2.5%	low	1.8%	64.3%	2026-07-07
12	KG	2.4%	low	1.0%	74.0%	2026-07-07
13	VE	2.1%	low	1.8%	50.1%	2026-07-07
14	BY	1.4%	low	1.0%	55.1%	2026-07-07
15	ET	1.4%	low	1.0%	59.0%	2026-07-07
16	NG	1.4%	low	1.0%	53.8%	2026-07-07
17	TM	1.2%	low	1.0%	44.8%	2025-09-26
18	GN	0.5%	low	1.0%	35.6%	2026-07-07
19	UG	0.5%	low	0.1%	86.6%	2026-07-07
20	CU	0.4%	low	0.1%	76.3%	2026-07-07
21	TR	0.4%	low	5.1%	1.6%	2026-07-07
22	AE	0.3%	low	0.1%	73.4%	2026-07-07
23	AZ	0.3%	low	0.1%	72.1%	2026-07-07
24	BF	0.3%	low	0.1%	74.5%	2026-07-07
25	BH	0.3%	low	0.1%	69.4%	2026-07-07
26	CN	0.3%	low	1.0%	8.9%	2026-07-07
27	EG	0.3%	low	0.1%	43.5%	2026-07-07
28	ES	0.3%	low	0.1%	40.8%	2026-07-07
29	IE	0.3%	low	0.1%	48.2%	2026-07-07
30	KE	0.3%	low	1.8%	4.8%	2026-07-07

1MMelevated
54.2%
Country risk
41.6%
OONI signal
84.8%
As of
2026-07-07
2PKelevated
27.6%
Country risk
17.5%
OONI signal
81.6%
As of
2026-07-07
3RUwatch
9.1%
Country risk
17.5%
OONI signal
48.1%
As of
2026-07-07
4TZwatch
9.1%
Country risk
6.7%
OONI signal
82.6%
As of
2026-07-07
5INwatch
6.0%
Country risk
41.6%
OONI signal
9.2%
As of
2026-07-07
6IQwatch
6.0%
Country risk
5.5%
OONI signal
70.1%
As of
2026-07-07
7IRwatch
6.0%
Country risk
5.5%
OONI signal
59.2%
As of
2026-07-07
8JOwatch
6.0%
Country risk
5.1%
OONI signal
52.5%
As of
2026-07-07
9SDwatch
6.0%
Country risk
1.8%
OONI signal
82.6%
As of
2026-06-19
10OMlow
5.9%
Country risk
1.0%
OONI signal
85.0%
As of
2026-07-04
11SYlow
2.5%
Country risk
1.8%
OONI signal
64.3%
As of
2026-07-07
12KGlow
2.4%
Country risk
1.0%
OONI signal
74.0%
As of
2026-07-07
13VElow
2.1%
Country risk
1.8%
OONI signal
50.1%
As of
2026-07-07
14BYlow
1.4%
Country risk
1.0%
OONI signal
55.1%
As of
2026-07-07
15ETlow
1.4%
Country risk
1.0%
OONI signal
59.0%
As of
2026-07-07
16NGlow
1.4%
Country risk
1.0%
OONI signal
53.8%
As of
2026-07-07
17TMlow
1.2%
Country risk
1.0%
OONI signal
44.8%
As of
2025-09-26
18GNlow
0.5%
Country risk
1.0%
OONI signal
35.6%
As of
2026-07-07
19UGlow
0.5%
Country risk
0.1%
OONI signal
86.6%
As of
2026-07-07
20CUlow
0.4%
Country risk
0.1%
OONI signal
76.3%
As of
2026-07-07
21TRlow
0.4%
Country risk
5.1%
OONI signal
1.6%
As of
2026-07-07
22AElow
0.3%
Country risk
0.1%
OONI signal
73.4%
As of
2026-07-07
23AZlow
0.3%
Country risk
0.1%
OONI signal
72.1%
As of
2026-07-07
24BFlow
0.3%
Country risk
0.1%
OONI signal
74.5%
As of
2026-07-07
25BHlow
0.3%
Country risk
0.1%
OONI signal
69.4%
As of
2026-07-07
26CNlow
0.3%
Country risk
1.0%
OONI signal
8.9%
As of
2026-07-07
27EGlow
0.3%
Country risk
0.1%
OONI signal
43.5%
As of
2026-07-07
28ESlow
0.3%
Country risk
0.1%
OONI signal
40.8%
As of
2026-07-07
29IElow
0.3%
Country risk
0.1%
OONI signal
48.2%
As of
2026-07-07
30KElow
0.3%
Country risk
1.8%
OONI signal
4.8%
As of
2026-07-07

API: GET https://api.voidly.ai/v1/shutdown-risk/{country}

Get notified when a country crosses elevated

Every band-crossing (low → watch, watch → elevated) is published as an RSS event the moment it’s detected. Journalists, activists, VPN providers, and AI agents can subscribe and react in real time — no polling needed.

Latest band crossings

19 predictions archived, 19 crossings detected to date

IRwatchfrom low → watchp = 12.5%on 2026-07-05view JSON →
INelevatedfrom watch → elevatedp = 15.7%on 2026-07-05view JSON →
SYwatchfrom low → watchp = 9.6%on 2026-07-03view JSON →
RUwatchfrom low → watchp = 10.9%on 2026-07-03view JSON →
JOwatchfrom low → watchp = 6.4%on 2026-07-01view JSON →

Live from /v1/shutdown-risk/alerts. Sorted newest-first. Each row links to the per-country endpoint with its full validation embedded.

RSS 2.0 feed

/v1/shutdown-risk/feed.rss →

Standards-compliant RSS. Add to Feedly, Inoreader, NetNewsWire, or any RSS reader. One item per country crossing, with the probability, model version, and a permalink to the live per-country endpoint.

JSON event log

/v1/shutdown-risk/alerts →

Same alerts, machine-readable. Poll from a script, an AI agent, or pipe into your own monitoring. Each event includes band, previous band, probability, KeepItOn evidence, and a permalink.

Push delivery (webhook)

One-click webhook subscribe — no account needed

#subscribe

Paste an https:// webhook URL. We send a test ping first — if it responds 2xx, we wire it up. Every band crossing for your selection is delivered as JSON via POST (same payload shape as /v1/shutdown-risk/alerts). The webhook URL is the secret; hold onto the unsubscribe link we return to revoke later.

Cooldown: 14 days per (country, band) suppresses repeat alerts. Computed daily by compute-shutdown-risk-alerts.py from the same prediction archive the accountability loop uses.

Live accountability

running_accountability tracks every archived shutdown_risk prediction (one per country per day, from archive-shutdown-risk-history.py) against the actual KeepItOn-documented shutdowns that followed in the 7 days after. early days have small sample size; validated_backtest contains the v4 trainer's static AUC-0.90 result (full-panel).

1498

Predictions archived

Flagged elevated

True positives so far

2025-09-26

Tracking since

The loop has just started — precision and recall stabilise once a few weeks of outcomes have landed.

API: GET https://api.voidly.ai/v1/shutdown-risk/accountability

Integrate it

The endpoints are free and CORS-open. Copy-paste to start.

curl — country risk

curl -s https://api.voidly.ai/v1/shutdown-risk/IN | jq .risk
# -> { "probability": 0.46, "risk_band": "elevated",
#      "components": { "v5_prob": ..., "sr_prob": ..., "hf_prob": ... },
#      "kio_hist_5y": 64, "kio_recent_7d": 0, "as_of_date": "..." }

Python — pull leaderboard daily

import requests
data = requests.get("https://api.voidly.ai/v1/shutdown-risk/leaderboard").json()
for cc, rec in sorted(
    data["countries"].items(),
    key=lambda kv: -kv[1]["probability"],
)[:10]:
    print(f"{cc}  {rec['probability']:.0%}  {rec['risk_band']}")

Node — subscribe to the RSS alerts feed

import Parser from "rss-parser";
const feed = await new Parser().parseURL(
  "https://api.voidly.ai/v1/shutdown-risk/feed.rss",
);
for (const item of feed.items) {
  console.log(item.pubDate, item.title);     // e.g. "RU: shutdown risk ELEVATED (p=22.8%)"
}

Voidly pushes to your webhook (no polling, no account)

# Register once. Every band-crossing for Iran (or any country) is POSTed
# to your endpoint in real time via the worker's scheduled fanout.
# Anonymous — the webhook URL is the secret. We send a test ping first;
# if it doesn't 2xx, nothing is stored.
curl -X POST https://api.voidly.ai/v1/shutdown-risk/subscribe \
  -H "Content-Type: application/json" \
  -d '{
    "webhook_url": "https://your-server.com/voidly-hooks",
    "country_code": "IR",
    "min_severity": "high",
    "event_types": ["shutdown_risk_elevated", "shutdown_risk_watch"]
  }'
# Response includes an unsubscribe_url you can save to revoke later.

# Payload your endpoint receives on a real crossing:
# { "event": "shutdown_risk.elevated", "country_code": "IR",
#   "date": "...", "probability": 0.46, "band": "elevated",
#   "from_band": "watch", "alert_id": "...",
#   "report_url": "https://api.voidly.ai/v1/shutdown-risk/IR" }

Verify your wiring without waiting for a real crossing

# Fire one synthetic shutdown_risk.elevated payload to your webhook
# right now. country_code is "XX" and test:true so downstream can
# sandbox it. Returns upstream HTTP status + latency.
curl -X POST https://api.voidly.ai/v1/shutdown-risk/test-alert \
  -H "Content-Type: application/json" \
  -d '{
    "webhook_url": "https://your-server.com/voidly-hooks",
    "band": "elevated"
  }'

Verify the HMAC signature (Python receiver)

# Every shutdown_risk webhook delivery (the test ping, the test-alert
# sample, and real band crossings) carries:
#   X-Voidly-Signature: sha256=<hmac-sha256(body, WEBHOOK_SECRET)>
#   X-Voidly-Event:     shutdown_risk.{connected|watch|elevated}
#
# Same WEBHOOK_SECRET signs all three — verify once, accept all of them.
# Contact us to receive your subscriber-specific secret; until then,
# subscriptions ship with a per-deployment shared default.
import hmac, hashlib
from flask import Flask, request, abort

app = Flask(__name__)
SECRET = b"your-webhook-secret"  # from /v1/shutdown-risk/subscribe response

@app.post("/voidly-hooks")
def hook():
    sig = request.headers.get("X-Voidly-Signature", "")
    expected = "sha256=" + hmac.new(SECRET, request.data, hashlib.sha256).hexdigest()
    if not hmac.compare_digest(sig, expected):
        abort(401, "bad signature")
    evt = request.json
    print(evt["event"], evt["country_code"], evt["band"], evt["probability"])
    return ("", 204)

Slack — pipe alerts into a channel

# In your cron / GitHub Action: poll the JSON event log and post new
# items to a Slack incoming webhook.
ALERTS=$(curl -s https://api.voidly.ai/v1/shutdown-risk/alerts | jq -c '.alerts[0]')
COUNTRY=$(echo "$ALERTS" | jq -r .country)
BAND=$(echo "$ALERTS" | jq -r .band)
curl -X POST -H 'Content-type: application/json' \
  --data "{\"text\":\"⚠️ Shutdown risk \\`$BAND\\` in $COUNTRY\"}" \
  $SLACK_WEBHOOK_URL

Honest caveats

·FORWARD-VALIDATED (the real product): in a strict past->future temporal holdout the CROSS-COUNTRY ranking holds — full-panel AUC ~0.87, country-ranking AUC ~0.90, 7.8x PR-lift over base rate. The model reliably tells you WHICH countries are heading into danger, forward in time. BUT the within-country median AUC is CROSS-SECTIONAL (full panel, in-sample isotonic) — it measures day-ranking when the model has seen the whole period, NOT forward skill. In a strict past->future temporal holdout, within-country timing AUC is ~0.38 (below chance), for BOTH v5 and v9, driven by non-stationarity (shutdowns cluster in time). USE THIS MODEL FOR *WHICH* COUNTRY IS AT RISK (cross-country full-panel AUC 0.90, which IS forward-meaningful), NOT as a within-country *WHEN* oracle.
·v9 replaces v5's multiplicative product with a logit-additive blend (alpha=0.6 on structural risk). The product collapsed to ~0 when honest_forecast was near-zero for a country, annihilating a good structural signal; the logit blend cannot zero out either factor.
·Within-country median AUC 0.715 vs v5 0.729. Improvement confirmed by leave-one-country-out alpha selection (all 23 folds chose alpha=0.60; held-out median +0.0095, mean +0.0131, 18/23 countries up).
·Still country-level; sub-national shutdowns (~half of KeepItOn events) remain harder. Low-N countries are sample-size limited, not model-limited.
·Same two honestly-validated base models as v5 (v4 structural + honest_forecast_v1). No new external features (unlike the gated v7 CF-Radar and v8 GDELT attempts) — purely a better combiner.
·WHAT 'low' MEANS HERE: this ranks the risk of a FULL national internet SHUTDOWN (KeepItOn-style total blackout), NOT general censorship. A heavy-censorship country (e.g. AZ, AE, EG) correctly reads 'low' if it rarely does FULL shutdowns -- it can still be among the most censored (see /censorship-index and /v1/atlas). The blend weights the shutdown-validated structural model (v4) 0.6 over honest_forecast 0.4 BY DESIGN: honest_forecast is a broad censorship-anomaly signal that is noisy for shutdowns specifically (it scores e.g. Spain ~0.78), and the 0.6 structural weight was leave-one-country-out validated. So a country with high honest_forecast but low structural risk is correctly down-weighted here.

For the related but distinct OONI-anomaly forecast (different question), see /v1/forecast/honest/{country}.