Trace every
conclusion.

The Global Censorship Index combines measurements from multiple networks into a country score, with classification and evidence processing between collection and publication. A score summarizes measured conditions; it does not observe every network or establish the cause of every anomaly.

The published v2.2 description says collection runs continuously and new blocks can appear within hours. Its limitations allow up to 24 hours for events to reach scores. Treat those as pipeline descriptions; use actual source timestamps when citing a result.

Probe coverage limits

The reference’s configured network footprint is not a live active-node count. The described deployments include the US, UK, Germany, Japan, Singapore, India, Brazil and South Africa. They help answer whether a domain is reachable from outside a country, and do not establish inside-country reachability. The validated source response above supplies the available node counts and region list.

The reference identifies limited inside-country coverage in IR, CN, RU, VE, EG, PK, MM, SY, TR, SD, KP, BY, UZ, TM, SA, VN, TH, LB and AZ. Inside-country evidence relies on volunteer OONI measurements and remote CensoredPlanet DNS/HTTP probing. Community operators can see the probe program.

Request checked: 2026-09-24T00:51:39.945Z. A page check is not a measurement timestamp.

No single network captures the complete picture. OONI provides active probing with geographic gaps. CensoredPlanet offers remote measurement with limited ground truth. IODA detects outages rather than selective blocking.

The v2.2 pipeline describes Voidly’s network testing VPN accessibility and censorship patterns every five minutes, then correlating those measurements with OONI, CensoredPlanet and IODA. Correlation produces structured incident records and linked evidence; an anomaly by itself remains ambiguous.

These reference metrics are dated. The bundled metric snapshot was generated 2026-08-27T05:29:58.087Z. The source’s May 2026 experiments and older forecast panels are identified separately below. Open the model registry preview for currently returned versions, source notes and dates.

Classifier v3.3 · bundled snapshot

Gradient boosting trained on 4,237 labeled samples, including 1,116 positive samples across 131 countries. The reference describes aggregate-data training with no raw user data. The live classifier record also explains that many positive labels are network disruptions; read its label-composition caveat before treating F1 as confirmed-censorship accuracy.

Algorithm

GradientBoosting

LOCO mean F1

0.71

Mean F1 for countries with n≥30

0.63 across 61 countries

LOCO median F1

0.87; inflated by 46 of 127 countries at a perfect 1.0 on very small samples

Stratified five-fold F1 / AUC

0.73 / 0.90; sanity checks, not deployment performance

Hard, high-volume countries

CN 0.29 · BY 0.21 · AZ 0.11

Training samples

4,237 (1,116 positive) across 131 countries

Reference retraining schedule

Weekly, Sundays at 02:00 UTC

The headline LOCO MEDIAN F1 (0.870) is dominated by many small-sample countries scoring a perfect 1.0 on a handful of days; the MEAN F1 (0.711) is the honest single number. Censorship-heavy, high-volume countries score materially lower (CN ~0.29 on n=95, BY ~0.21 on n=84, AZ ~0.11 on n=65). Cite the mean — or the specific per-country number — for hard countries, not the median. The model is CLEAN (no label leakage); this is a distribution caveat, not an accuracy retraction.

The earlier v2 reported 99.8% F1 and 1.000 ROC AUC, and was retired on May 21, 2026 after an audit found that the label-derived country_risk_tier feature contributed 85% of importance. Read the leakage audit.

Classifier feature importance · bundled reference

v3 removed the country-tier shortcut. The original v3 account describes roughly 73% importance in its top three signals and no single dominating feature. The following stored importance list belongs to the later v3.3 reference, with 13 base features plus three regime-similarity-weighted contagion features; do not mix the two generations.

anomaly_rate

22.1%

month

20.2%

measurement_count

17.0%

neighbor_max_anomaly_7d

8.9%

neighbor_incident_count_7d

7.7%

neighbor_block_rate_7d

7.6%

rate_count_interaction

5.6%

v3.3’s reference records stratified F1 0.729, LOCO mean F1 0.711 and inflated median F1 0.870. Inspect current feature importance.

Legacy Sentinel forecast · bundled snapshot

XGBoost with isotonic calibration is a separate model from both classifier v3.3 and honest_forecast_v1. Its legacy target_7day sliding-window label measures the prevailing regime; these scores must not be attached to the newer OONI-event forecast.

Training-holdout LOCO median AUC / F1

0.88 / 0.51

Snapshot status

DEGRADED — rolling 30d precision 0.172 < 0.6

Rolling 30-day precision / recall

0.17 / 32%

Rolling Brier / calibration MAE

0.28 / 0.29

Evaluated country-days

898

Training records

14.6K

Reference retraining schedule

Weekly, Sundays at 02:00 UTC

The public accuracy endpoint describes daily evaluation over a rolling 30-day window. The values above are the dated bundle, not a fresh endpoint fetch.

The published index uses a 0–100 scale: 0 denotes complete freedom in the scale definition, and 100 denotes total censorship. These are descriptive score bands, not forecast probabilities or the separate data-quality confidence bands.

0–10 · Free

Minimal or no censorship

11–25 · Low

Limited content restrictions

26–45 · Medium

Significant restrictions on some platforms

46–70 · High

Widespread blocking of platforms and news

71–100 · Severe

Pervasive censorship / isolated internet

• National averages do not capture regional variation.
• VPN detection is underreported in highly restricted environments.
• Sample sizes vary by country and affect confidence.
• Real-time events may take up to 24 hours to reach the scores.
• Content filtering and throttling are harder to detect than blocking.
• Self-censorship and legal restrictions are not measured.

The reference describes country-score intervals reflecting measurement certainty. Wider intervals indicate less data or greater variability. The examples below are illustrative only; they are neither current countries nor a forecast validation result.

Open actual index scores. A measurement-confidence interval, a model score, a probability and an empirical forecast-coverage statistic answer different questions.

The published workflow compares scores with external benchmarks, known events and three holdout approaches. The Freedom House Freedom on the Net comparison is a self-reported correlation of r=0.87. The reference gives Iranian shutdowns matching score spikes as an example of event checking.

Stratified k-fold

Random splits provide a sanity check but can leak temporal and country patterns. Do not cite them as deployment performance.

Leave-one-country-out

Train on all other countries, evaluate on the held-out country. This tests geographic transfer, not future-time generalization.

Time-based

Train before T, test after T. This tests temporal transfer and can expose failure on novel events.

Classifier and forecast results remain separated in the model section above. The accuracy source explicitly warns that stratified AUC can overstate the time-based split by 47.9 percentage points, and recommends LOCO or populated rolling-production results. Even a LOCO median needs the classifier’s small-sample caveat; it is not universally the best headline. Read the replication guidance.

1. OONI
2. Ingestion
3. Feature engineering
4. ML scoring
5. Index update

Published v2.2 schedule, rather than a guarantee that the last run completed:

Classifier retraining

Weekly, Sundays at 02:00 UTC

Publication

Daily at 03:00 UTC

Score latency

Approximately six hours for aggregates and five minutes for probes

Raw ingestion

Every five minutes for probes; every six hours for OONI, IODA and CensoredPlanet

The reference describes KS-test drift detection on a rolling holdout triggering extra retraining, a D8 promotion gate covering G1–G7 plus calibration, thresholds for AUC/recall/coverage, and a post-promotion watchdog that rolls back regressions. It describes automated publication without a human in that loop. The existing operational page carries the counters and source; the pipeline diagram here does not verify a particular run.

Cite the methodology page

Cite the index dataset

Voidly Research. (2026). Global Censorship Index. https://voidly.ai/censorship-index

@misc{voidly_censorship_index,
  author = {Voidly Research},
  title = {Global Censorship Index},
  year = {2026},
  url = {https://voidly.ai/censorship-index}
}

Voidly-original incidents, scores and annotations are CC BY 4.0, permitting reuse with attribution. Raw upstream measurements keep their own licenses. As the data page specifies, OONI raw data and Citizen Lab test lists use CC BY-NC-SA 4.0; redistribution of those subsets must honor those terms. Source-specific data and licensing.

• Full index JSON with schema.org markup
• Index CSV for Sheets, Excel and Pandas
• Machine-readable methodology
• Detection, prediction and realtime API documentation

The machine-readable methodology retains some legacy model fields. Use the versioned model source for its features, target and evaluation; do not infer one common model from the general method document.

Research

research@voidly.ai

Partnerships

partnerships@voidly.ai

General

team@voidly.ai

View the index · Live alerts