How we read a story

01Independent voices

"45 outlets reported it" is the number everyone shows, and it is the number that lies. We collapse every known dependence between outlets and count what remains. Two outlets are one voice when they carry the same text, when the same owner controls both, when both credit the same wire agency for the text, or when their "own words" reports share two or more identical quoted sentences, which means one briefing. Unknown dependence is never counted as independence; it lowers the confidence shown beside the count. Independence is voices divided by outlets.

The ladder beneath the count: one voice; one voice, many copies; three or more independent voices; a primary source seen.

02Origin and source grade

The origin is the earliest report we can establish, using explicit credits in the text, links from later reports to earlier ones, the first appearance of each quoted sentence, the outlet's own timestamp weighted by how reliable that outlet's timestamps have proven, and our own clock. When two candidates fall within the same minute we say "tie" rather than pick.

The root of a story is graded: P1 a primary document (a filing, statement, dataset or official page we reached); P2 a named person or institution on the record; P3 a wire agency's text; P4 an outlet's own reporting with a byline; P5 an anonymous source ("sources say"); P6 unverified, because the page could not be read or carries no byline, credit or primary link.

03Claims by claimant

Figures in headlines and ledes are tied to who said them. The same claimant giving a new figure is marked changed. Different claimants giving different figures are marked disputed. We never resolve a dispute; we show it.

04How it spread

A text family is the set of reports carrying the same words. We measure the spread of wire-copy families across our own scan and call a family pushed only when 60% or more of its carriers belong to one owner, its burst is tighter than the fifth percentile of normal wire copy, and it carries no agency credit. Six push signals are counted and shown with their evidence: a PR-wire or sponsored origin; a burst tighter than wire; one owner behind most carriers; no reachable source with many copies; paid representation on the public record (FARA) for a country that is party to the story, shown as context; and a habitual network of outlets that co-publish each other's texts.

05Every side

For every country that is party to a story we read that country's own press in its own language and show what it says, with each outlet's owner beside it. When a side has not spoken we say so and how many of its outlets we read. The Unheard list shows stories carried by three or more outlets of one country, from at least two owners, that no outlet elsewhere picked up.

06Owners

Every owner label comes from a registry (Wikidata, RSF's Media Ownership Monitor, the FCC, company registers) and is marked verified only when two independent sources agree. One-source rows say "owner not verified".

07Edits

Pages are re-read at one, six, twenty-four and seventy-two hours. When the headline or lede changes, both versions are kept and shown.

08Verification, signing and sealing

Before a card is published a second, independent pass re-opens the evidence and removes anything that no longer checks out. The card is then signed and its version sealed into an append-only hash chain that is timestamped on the Bitcoin blockchain. A dossier can be checked years later even if its sources vanish; archived copies of evidence links are kept for the same reason.

09Accuracy, measured daily

The engine tests itself every day: consistency across every dossier, planted cases with known answers, and a human-audited sample when one exists. The scores and their date are published at /api/accuracy.json and summarised here. Targets: grouping precision 95%, origin accuracy 90%, copy precision 98%, planted recall 100%. A miss is published, not hidden.

11Versions

Every number carries the version of the method that produced it. Method changes get a new version number and a dated note here; old numbers are never recomputed under new rules.

• 4 October 2026: a story's record is sealed into the chain only when it materially changes: a report joins or leaves, a claim is added or changes, the independent-voices rung moves, the origin changes, or an edit is detected. Counts and timestamps alone no longer create a new sealed version. The chain is unbroken across the change.
• 4 October 2026: when the earliest report reaches us only as a Google News link, we look for the same outlet's own page: first by the exact headline, then by a near-identical one that must pass three tests together (most words shared, at least four words in the same order, and no second page nearly as close), using the outlet's own feed and its news sitemap. A page found this way is marked as matched by title. When nothing passes, the origin stays unverified and says so.

Current method: print-v3