How Zaviye reads, and what it misses

502 feeds. 269 are English, 53 Arabic, 32 Persian, 24 Spanish, 20 Turkish, 20 Kurdish in Sorani and 3 in Kurmanji, 15 Hebrew, and 66 spread over 18 other languages.

A feed gives us a headline and a line of text. In the last week, 95% of the 76,001 articles we collected came with fewer than 200 characters of text, and the median was 98. We read headlines. We do not read the articles.

381 of the 502 feeds reach us through Google News search, not from the outlet. Google passes on the correct time but shows the item hours after the outlet posted it: in October 2026 the median was about 3 hours for Milliyet and El Tiempo, and about 6 for Hurriyet and Clarin. Only 76 of our 358 busiest sources (those with 60 or more articles in a month) stamp articles within an hour of when we fetch them; they supply 45% of the articles, and most come straight from the outlet.

We do not read television, radio, video, print that is not online, or anything behind a paywall past its headline.

The six languages we list on the front page are not equal in the data. Of the 4,142 stories with three or more articles in the last 30 days, 65.1% contain English, 17.2% Turkish, 5.0% Arabic, 2.4% Persian, 0.7% Kurdish and 0.1% Hebrew (6 stories).

So an empty Hebrew or Kurdish bar on a story page is the usual case, not a finding. A missing bar is not a silence.

The list of feeds also keeps changing. In April it held 423 feeds. On 1 and 2 June we added 30 Arabic and 11 Kurdish feeds, so anything we counted before then is not comparable with anything after.

By a formula, not by a person's pick. A story scores the number of different outlets that carried it plus a quarter of its article count. That score halves every 15 hours after the story's newest article. It is multiplied by 1 plus 2.5 times the share of the story's outlets that sit in the region we cover. If a reader has set a country, a locality score for that country can add up to half again.

The top 200 stories are shown. The top slot goes to the best-ranked story whose outlets are at least half regional, if there is one. A story whose headline is not yet in English waits until it is.

Articles are grouped by the words in their headlines (35% overlap) and the English headline is written afterwards. The same event can end up as two stories, one per language.

We found out how often by checking our own work. Of the 27 silences we published in Wires 1 to 9, 10 were an event our grouping had split in two: one cluster in Turkish, the same event in English or Urdu, each calling the other silent. 4 were things that happen every day in every language (a death in Gaza), and 3 came from a single outlet. We changed the rule (section 6).

When ten outlets run one agency report, that is one report. On 18 October 2026, the multi-outlet stories held 4,395 articles and 3,644 independent reports: about one in six was a copy. In the Fort Hood execution story it was nearly half (79 articles, 43 independent).

We find copies by near-identical headlines, shared figures and agency credit lines. We miss translations and rewrites, so real independence is lower than these counts. Two feeds of one publisher count as one outlet.

A story is called a silence only if all of this holds:

• at least three outlets carried it;
• no story in any other language matches it, checked against everything we collected that week, not only what reached the front page;
• the chance that every other language would stay silent by luck is 10% or less, given how often each language carries stories of that size.

Replayed over 12 past weeks, this gives 23 silences where the old rule gave 36. 3 weeks produce none, and a week with none publishes none.

What survives is almost all Turkish. Across 13 weekly files, stories carried by 10 or more outlets, all of them Turkish, happened 62 times; luck predicts about 12. English-only stories sit at luck's level (356 against 310), so “only English carried it” is not a silence.

For Persian, Kurdish, Arabic and Hebrew the test finds nothing, and that is not good news. In those weekly files not one story with ten or more outlets was carried by any of them alone, where luck predicts about 6 for Persian, 5 for Arabic, 1 for Kurdish and 0 for Hebrew. These languages rarely appear together with others in large stories, so the test cannot see a silence in them. Fixing that needs better joining of stories across languages, not a looser rule.

For 81 stories carried by both English and Turkish outlets whose clocks we trust, the English outlet published first in 79% of cases. Volume alone would give 67%. The median gap was 3.6 hours.

For every other pair of languages we measured no reliable difference, and for Persian, Kurdish and Hebrew we have too few stories with trustworthy times to say anything. Most of our busy sources cannot be timed this way, because of the Google delay described in section 1.

The English headlines, the short summaries and the translations into Persian, Arabic, Turkish, Kurdish, Hebrew and Spanish are written by language models, some running on our own server and some through a cloud provider. The Wire and the Dossier are drafted by a larger model. No person reads or edits any of this before it publishes. Models are replaced when they are retired or beaten.

Everything is written in English first and translated afterwards. Nuance that English flattens is flattened for every language.

• A story is split in two, or two stories are joined.
• A copy is counted as a second report.
• A model mistranslates, most likely in Kurdish, the language here with the least training text.
• A headline misleads and the article says otherwise. We never read the article.
• A feed breaks for days before we notice.

If you find one, write to [email protected].