Sourcing
We build each article from licensed and openly-licensed sources — news wires, RSS feeds, and public records — rather than scraping the open web. For any given event we gather coverage across many outlets and languages, so no single publication's framing decides the story.
Claims, not prose
Before writing, we break the source coverage into discrete, checkable claims — extracted on our own hardware by a local model, never sent to a third-party API. Claims are clustered so that the same fact reported by ten outlets is recognised as one, with its supporting sources attached. What we store is this structured evidence, not copied text.
Cross-source synthesis
An article is composed from the clustered claims, weighing corroboration and noting disagreement. Where outlets conflict, we say so. Where a claim rests on a single source, we flag it. Every substantive statement links back to the sources behind it, listed at the foot of the piece.
Written under a named byline
The final piece is written by a large language model under a consistent, named persona — and edited to our standards. The bylines are personas, not people; how that works, and how we handle libel and corrections, is set out in Standards.
Multilingual by construction
Each article is produced in six languages. Because the underlying claims are language-neutral, the translations describe the same evidence — and our search works across all of them, so a term in any language finds the piece.
Linked to the record
Every article links to the underlying event — a living page that tracks the sources and how coverage evolved over time. It is the audit trail: you can always follow a story back to where it came from.
Spotted something wrong? Our corrections policy explains how we fix it.