Global censorship.
In the evidence.

Source citable category

894

Category records with linked evidence

807

Category records with multiple sources

438

citable_censorship = censorship/mixed incidents excluding suspected/draft; this is a type/status count. citable_with_evidence = the subset linked to >=1 existing evidence row outside the historical OONI 'normal' stored-label quarantine pending regrade; its URL may be search or dashboard context rather than an exact measurement. citable_multi_source = the subset whose incident.sources lists >=2 source labels; this count does not establish agreement or independence.

The source category, individual evidence links and multi-source corroboration are different properties. Open the incident before citing its evidence.

Country rows retrieved

130

Source-window days / measurements

7 / 390,258

derived: sum of countries[].samples (OONI measurementCount) over the rolling 7-day live window. Not floored, not padded, not a constant.

Newest per-country ingestion stamp, not proof that every country is fresh. Each country row has ingestionAsOf; none is a measurement observation time. `dateModified` and `generated_at_utc` are serve times.

This index route does not read a model artifact or evaluation. Read the serving classifier and its LOCO mean F1 at /v1/classifier/info; check forecast evaluation separately at /v1/sentinel/accuracy and /v1/shutdown-risk/info.

{
  "scoringVersion": "2026.10-eb-k10-w7",
  "method": "ooni_anomaly_rate_7d_shrunk_empirical_bayes",
  "windowDays": 7,
  "windowBasis": "Seven days, from the upstream aggregation source (intelligence/src/services/ooni.ts windowStart). The deployed upstream service does not report its window, so this is the source value, not a live readback. Until 2026-10-19 this index was mislabelled as a 30-day window (method ooni_anomaly_rate_30d_shrunk_empirical_bayes, scoringVersion 2026.08-eb-k10); the scores were the same.",
  "previousIdentifiers": {
    "method": "ooni_anomaly_rate_30d_shrunk_empirical_bayes",
    "scoringVersion": "2026.08-eb-k10",
    "until": "2026-10-19"
  },
  "formula": "score = round(100 * anomalyRate * n / (n + 10))",
  "shrinkK": 10,
  "levelBasis": "level is derived from the unrounded shrunk rate (severe >= 0.7, high >= 0.5, medium >= 0.25, low >= 0.1, else free) — NOT from the rounded integer score. A country with samples: 0 is \"unmeasured\", never \"free\".",
  "levelValues": [
    "severe",
    "high",
    "medium",
    "low",
    "free",
    "unmeasured"
  ],
  "unmeasuredMeaning": "samples: 0 means OONI published NO measurements for this country in the 7-day window. That is the absence of evidence, not evidence of absence, so such rows are level \"unmeasured\" and are excluded from summary.freeCount. Before 2026-08-29 they were published as \"free\" — North Korea, Eritrea, the Bahamas and Nicaragua were all served as free countries on zero measurements.",
  "rankTieBreak": "shrunkRate desc, then samples desc, then code asc — a total order, identical in every distribution",
  "lowConfidenceBelowSamples": 30,
  "appliesTo": [
    "https://api.voidly.ai/data/censorship-index.json",
    "https://api.voidly.ai/data/censorship-index.csv"
  ],
  "changedAt": "2026-08-29",
  "changedFields": {
    "https://api.voidly.ai/data/censorship-index.json": {
      "unchanged": [
        "score",
        "samples",
        "country set"
      ],
      "changed": [
        "rank",
        "level"
      ],
      "detail": "score: 0 of 130 rows changed — byte-identical to what was served before. rank: 108 of 130 changed, 50 of them toward rank 1, entirely within groups whose score did not move (largest: SE 121 to 95, LT 86 to 61, SI 91 to 64, all at score 1-2; among rows scoring 10 or above only HN 9 to 8 and IR 8 to 9 swap). level: 4 of 130 changed, all of them zero-measurement rows moving from \"free\" to \"unmeasured\" — see unmeasuredMeaning."
    },
    "https://api.voidly.ai/data/censorship-index.csv": {
      "unchanged": [
        "samples",
        "country set",
        "columns 1-8 by name, order and meaning"
      ],
      "changed": [
        "rank",
        "score",
        "level",
        "4 columns appended"
      ],
      "detail": "score: 12 of 130 rows changed, every one DOWNWARD (largest TM 50 to 8). level: 3 changed downward (TM high to free, DZ and PK low to free) plus the 4 zero-measurement rows to \"unmeasured\". Any row taken from this file before 2026-08-29 should be re-fetched. Four columns were appended (raw_anomaly_rate, shrunk_rate, low_confidence, scoring_version); a reader that pins the column COUNT at 8 will reject the file, one that reads by position or header name will not."
    }
  },
  "changeNote": "Before 2026-08-29 the CSV distribution published the UNSHRUNK anomaly rate while the JSON published the shrunk one, and the two sorted differently. Turkmenistan was served as rank 1 / score 50 / high in the CSV on 2 measurements; its shrunk score is 8 / free. 12 scores, 3 levels and 124 ranks disagreed between the two files. Both are now computed once. NO SCORE OR LEVEL MOVED UPWARD IN EITHER FILE — verified against the full live 130-country corpus. BOTH distributions changed, not only the CSV: the CSV s scores and levels moved down, and the JSON s `rank` changed for 108 of 130 rows because ties on the rounded score used to be broken by upstream array order. See `changedFields` for the per-distribution breakdown. A consumer that stored `rank` from EITHER file should re-fetch."
}

The original report called this a composite of block rate, category breadth and consistency. The actual scoring definition above belongs to the retrieved index; its window and methods can differ from historical report text.

Collection documentation

• The original page describes roughly 1,100 distinct domains per day from a 5,154-domain pool across 130 countries. Its node count was a build-time constant, not a verified active roster.
• It describes OONI ingestion for eight test types across 130 countries every six hours; CensoredPlanet DNS/HTTP/S data across 50 countries; and continuous IODA ASN-level outage alerts.
• These are documented source/configuration descriptions, not fresh coverage or delivery checks. Read the current probe roster and source dates before quoting them.

Classifier v3.3 reference

GradientBoosting: LOCO mean F1 0.71 across 127 countries; 0.63 on the 61 with n≥30. The 0.87 median is inflated by 46 small-sample countries at a perfect 1.0. Training used 4,237 labelled samples, including 1,116 positive samples across 131 countries.

Shutdown-risk reference

The original shutdown-risk evaluation reports forward country-ranking AUC 0.90 against Access Now KeepItOn shutdowns. It ranks which countries; within-country day timing AUC is 0.33, below chance. These are not the distinct Atlas honest_forecast_v1 event model's scores.

The original page describes cross-source correlation and an incident-generation schedule every 30 minutes. A documented schedule is not proof that each run succeeded, and correlation is not independent confirmation of every incident.