Voidly

Voidly Research Reading the map 02

The denominator behind a censorship map

How to read Atlas country scores, sample counts, and missing data without mistaking an anomaly signal for a census of internet freedom.

Index read October 4, 2026; newest country ingestion 15:21 UTC

Data snapshot: October 4, 2026. Newest per-country ingestion was 15:21 UTC; individual country rows may be older. API response read at 18:13 UTC.

A map of internet censorship invites a quick reading: darker here, lighter there. That view can be useful, but it leaves out the number that should come first: how many measurements produced each color?

Voidly Atlas's country index listed 130 country entries when we read it on October 4, 2026. It reported 455,527 upstream OONI samples in its rolling 30-day window. The reported sample counts are unevenly distributed across the index. The index is a way to inspect a measured signal at a particular time. To use it responsibly, a reader needs the sample count, the age of the source data, and a route from a country score to the underlying evidence.

What the score measures

The index's scoring method starts with the share of available OONI measurements marked anomalous in each country during the 30-day window. It then reduces the influence of small samples before rounding to a 0–100 score. The published method describes this as an anomaly rate with empirical-Bayes shrinkage. In plain terms, a country with a handful of observations should not receive the same apparent certainty as one with thousands.

An anomaly is a reason to investigate. OONI's own guide explains that an anomalous test can arise from potential interference, but it can also be a false positive. Interpreting a suspected block requires looking at the test, the network, repeated observations and other explanations for the failure. A score built from anomaly rates therefore cannot tell us what share of a country's people experienced a block. It does not identify the actor behind a disruption.

That distinction affects how we write about the map. “The index scored this country at 32” is a statement about this dataset and method. “Thirty-two percent of the country was censored” would describe something the index never measured. The first sentence can be checked against the public response. The second would need entirely different evidence.

A similar score can rest on very different volumes

The October 4 snapshot gives a concrete example. Russia had a score of 36 from 45,085 OONI samples in the window. China had a score of 32 from 485 samples. Both appeared in the index's “medium” band, but the underlying measurement volumes differ by nearly two orders of magnitude. That band describes the measured anomaly signal; it cannot establish that access restrictions in China are moderate.

Coverage itself can be constrained. OONI documented blocking of its Probe app and backend services in China, along with a drop in measurements, and warns that running Probe can carry risks. Inference: with only 485 samples in this snapshot and these known participation constraints, China's score likely understates restrictions that the index cannot observe. The available data do not support a numeric correction or a countrywide prevalence estimate.

The distribution is uneven across the whole map. The ten countries with the most samples contributed 255,520 of the 455,527 observations, or 56.1% of the total. That calculation does not make the remaining country rows useless. It tells us to keep the denominator beside the color, especially when comparing places or looking for a change over time. A rise or fall in a score deserves a look at what was tested and how many observations entered the window.

The countries with the most samples in this snapshot include the United States, Russia, Spain and Germany. Their volumes do not rank the severity of censorship; they rank the amount of OONI data available to this particular calculation. A visitor should be able to see the difference without learning a statistical model first. Atlas exposes the sample count in each country row, so the check is available in the same response as the score.

Missing data needs its own label

Six of the 130 entries had zero OONI samples in this 30-day window: the Bahamas, Eritrea, North Korea, Laos, Nicaragua and Turkmenistan. Atlas labels them “unmeasured.” A zero-score color would make absence of observations look like a measured finding of open access. The unmeasured label avoids that mistake.

The label is narrow. It says this dataset had no OONI samples for those country rows in this window. It says nothing about whether someone elsewhere collected other measurements, or what residents experienced. It also means these six rows should not be counted alongside measured low-anomaly rows. The index's own change note says earlier distributions had represented some zero-sample rows as “free”; the current response separates them.

A useful map should make that gap visible rather than smoothing it away. The person using it can then decide whether to gather more observations, consult another source or leave a claim open. Each of those is more defensible than treating a blank spot as reassurance.

Move from a score to a specific record

A country score is a broad screening tool. An incident record answers a different question: what did the observatory associate with a dated event, and which source links support that association? For example, a Voidly record dated September 30, 2026 classifies a Venezuela incident as mixed connectivity and censorship. It lists 41 evidence links across IODA and Censored Planet. That record gives a reader discrete items to inspect; it does not establish a cause or the number of people affected.

The October 4 country index gave Venezuela a score of 22 from 17,833 OONI samples in its rolling window. The score and the incident should not be merged into a single claim. They cover different windows and sources, and each has its own limits. A careful account can say what the 30-day signal was, then separately cite the incident and its evidence links.

Before using any country score, we suggest four quick checks: read the index's newest-ingestion time and the country row's own freshness; note the sample count and 30-day window; identify whether the row is measured; and open a dated incident or upstream measurement when making a specific censorship claim. In the October 4 response, the newest per-country ingestion was 15:21 UTC. That timestamp does not date every country row. The later API serve time was not a fresher measurement.

The point of a public index is that these checks can be repeated. Readers can download the same response, see the same denominators and challenge an interpretation with a specific record. The map can direct attention; the evidence decides what the story can say.