How digesta works
Every number and every verdict on this site is produced automatically. This page explains how — so you can judge the method, not just the output.
Where the material comes from
digesta reads public news sources — RSS feeds, Telegram channels and X accounts — across dozens of countries and several languages. It does not employ reporters and does not publish original reporting. Everything you read here is derived from what other outlets published first.
How stories are grouped
Items about the same event are grouped into one story. Candidates are found by comparing meaning rather than wording, so the same event reported in Russian, English and Portuguese lands together. Borderline pairs — and anything arriving long after a story opened — are decided by a language model that is shown both publication dates, because a follow-up, a correction and a fresh incident can look nearly identical to a similarity score.
Two kinds of text are deliberately kept out of grouping: multi-topic digests («evening roundup» posts, which glue unrelated events together) and daily columns — horoscopes, weather, exchange rates. A column is not an event, and every outlet runs one every day.
How the summary is written
Each story gets one neutral summary, written by a language model from the source texts. The rule that shapes it is corroboration: the model counts how many distinct outlets state each candidate fact, drops what only one outlet says, and orders the rest by that count. Where outlets give different figures, the summary gives the range and names the outlets at its ends — unless the dates show the number was simply rising, which is an update rather than a disagreement.
The summary leads with the latest development, not the first one. This matters more than it sounds: a big story keeps arriving for hours, and a summary built only from its opening hours describes a world that has moved on.
What «confirmed by independent sources» means
Outlets are not independent voices by default. Several titles owned by one group, or all writing for one national media space, frequently carry the same claim because they carry each other. So facts are weighted by independence: correlated outlets share a vote, and a fact enters the story's core only if it clears a share of the story's genuinely independent voice.
A fact outside the core is marked not independently confirmed. That is not the same as false, and not the same as copied. It means the corroboration we can measure did not reach the threshold — which can happen to a true fact that only part of the coverage reported, and to a fact carried by several outlets that our measurement treats as one voice.
The consensus map
The circular chart on a story is a map of how its coverage diverges. Each dot is one outlet. Distance from the centre is how far that outlet's account sits from the common account; the direction encodes how it differs, so outlets framing a story the same way cluster in the same sector. Colour runs on the same scale as the distance. The map describes divergence in wording and emphasis — it is not a score of honesty.
Fact-checking
Some stories are additionally checked against official and primary sources — the ministry, the company, the court, the agency that would know. This runs only where a story is carried within a single media space, because a story covered independently across several countries has already been corroborated by that coverage itself.
Each citation is labelled by what it is: an official or primary source, or a media report about one. Verdicts are produced automatically and are not editorial judgments. Checks carry the story's own date, because official sites publish near-identical pages for every recurrence of a regular event, and an undated check will happily verify this year's story against last year's announcement.
The media rating
The public rating measures two different things and keeps them apart. Metrics of factual conduct — how often an outlet's claims are corroborated, how complete its account is, how it behaves on stories where coverage diverges — are what the ranking is built from. Presentation metrics, such as tonal slant, are shown but never ranked: a strong voice is a legitimate editorial choice, and ranking outlets by it would be ranking them by taste.
An outlet is scored only once we have seen enough of its stories to say anything at all; below that it appears without a rank rather than with a noisy one. The rating is a measurement over a moving window, not a verdict on an organisation.
Limits, stated plainly
- Everything here is automated. There is no editor deciding what is true.
- Summaries are written by a language model from source texts, and can be wrong in the ways language models are wrong.
- Corroboration measures agreement among the outlets we read. Agreement is not truth, and outlets we do not read cannot be counted.
- Which sources are in the pool shapes every number on this site. The pool is uneven across countries and grows over time.
- We do not link to the original articles: the product stores what outlets published, not addresses to send you to.
Corrections and contact
If something here is wrong, write to [email protected]. That address is also the channel for data-deletion requests described in the privacy policy.