How this is counted
The site asserts that two publications are about one event, and that some country said nothing about it. This page says on what grounds.
What a divergence is
Publications are collected from outlets' feeds, reduced to text and turned into numbers by a multilingual model — it places texts close in meaning near one another regardless of the language they are written in. Each new publication then looks for its nearest neighbour among those already processed in the last 72 hours. If they are close enough, it is one story; if not, a new one begins.
A divergence is what the site calls a story that outlets of more than one press bloc wrote about. Stories from a single bloc do not go anywhere — they are in the country sections.
How close counts as «enough»
There is not one threshold but several, and each was calibrated against the corpus rather than assigned.
| 0,94 | within one language |
| 0,88 | across languages — the model places translations further apart than retellings |
| +0,03 | a surcharge when both publications come from one outlet |
| 0,90 | a match against the outlet's own archive: this is not an event but a column |
The surcharge for a shared outlet is needed because a newsroom resembles itself: one day's weather forecast is almost indistinguishable from the next day's. The recurrence check catches the same thing from the other side — an exchange-rate table, a broadcast announcement, a weekly column. Such a publication gets a story of its own, but cannot join anyone else's.
Where the site errs on the side of caution
The thresholds were checked against 112 hand-labelled pairs of publications. Those show that within a language the best result comes at 0,93, while the site runs at 0,94 — that is, it merges less often than it could.
That is deliberate. Wrongly merged publications look like a divergence that never happened — which is exactly what the site promises not to show. A missed match makes it poorer; a false one makes it untruthful.
What the machine says and what a person says
Outlets' headlines are translated by a model, and every machine translation is labelled — both on the story page and on the card. The translation is needed because Azerbaijani, Turkish and Georgian outlets do not publish in Russian at all: without it half the voices on the Russian site would be mute.
Everything else is written by an editor. The story's headline, the line saying what happened, and the conclusion about whether outlets agreed or diverged are the site's own voice, and they are not entrusted to a machine. A story does not reach the site without an editor's confirmation.
What the site does not publish
From each publication the site shows the headline, an excerpt no longer than 200 characters and a link to the outlet. The full text is stored only for analysis and is never handed out: the cut is made on the server, and the full text never reaches the page.
Read the article where it was written. The product here is the comparison, not the articles themselves.
If something here needs correcting or taking down — how to do it.
What the site cannot see
A bloc's silence means exactly one thing: among the outlets we read in that country, nobody wrote about the event. It does not mean the country was silent — beyond the list of outlets the site is blind.
That is why the list of outlets we read is open: who we read. A country with one feed in the registry is silent in a different way than a country with six.