How Reed reports

These policies govern every article we publish. They're written down here so you can hold us to them.

When we publish, and when we don't

Reed publishes an article only when at least one of these is true for the underlying event:

Two or more independent sources have reported it.
One government, official-agency, or institutional source has reported it — court records, police reports, parliamentary proceedings, EU institutions, statistics agencies, regulator filings, and the like.

A single non-official source is not enough. The pipeline drops the event before any draft is written. We'd rather miss a story we could have had than publish one based on a thin record.

Sources count as independent when they're separate publications under separate editorial control. Two newspapers running the same wire-service story counts as one source. Public service broadcasters (SVT, BBC, NRK, DR) are treated as established media; they need a second source.

The exception in the other direction: once an event belongs to an evolving story, every update gets a deeper review with web research and full claim-level corroboration — typically 15+ sources — regardless of how many publications cover the latest beat.

How confident we are about each fact

Every claim Reed publishes carries a confidence treatment that determines how it appears in the article text. The treatment is decided by code based on the source mix behind the claim, not by the model writing the prose.

Confidence	What it means	How it appears
High	Two or more distinct source categories agree.	Three people were injured.
Medium	Two or more sources in the same category, or one official source.	According to police, three people were injured.
Low	Single non-official source.	Aftonbladet reports that three people were injured.
Disputed	Sources contradict.	Police say three were injured. Witnesses say five.

You can see the confidence treatment for every sentence by clicking transparency on any Reed article.

Every source has a name and a link

Reed does not use unnamed sources. Every claim ties to at least one published source — a press release, a wire-service report, a newspaper article, an official document — with a link the reader can open. If a fact can't be cited by name and URL, it doesn't make it into a Reed article.

This is a deliberate trade-off. Investigative journalism sometimes depends on confidential sources; we forgo that capability in exchange for verifiability. We can only report what the broader source ecosystem has already attributed publicly.

Source categories

We classify every source into one of five categories. The category affects how much weight a source carries in the confidence calculation above, and you can see the category for every source on each article's transparency view.

Government & official

Press releases, court records, parliamentary proceedings, official statements, regulator filings, statistics agencies.

Wire services

TT, Reuters, AP, AFP. The independent reporting infrastructure that newsrooms across the world rely on.

Public service media

SVT, SR, BBC, NRK, DR, and other public broadcasters with statutory editorial standards.

Established media

Major newspapers and broadcasters with editorial accountability — for example Dagens Nyheter, Svenska Dagbladet, The Guardian.

Other

Local outlets, specialized publications, and verified accounts. Used as supporting evidence; never as a sole source for a published claim.

A claim backed by two different categories gets the highest confidence. A claim backed by one official source gets medium confidence. Anything weaker is flagged in the article text itself rather than presented as fact.

When sources disagree, you'll know

Reed surfaces contradictions instead of choosing a side. When two sources we trust make conflicting claims about the same fact, the article presents both — "Police say X. The opposition says Y." — and the transparency view lists the contradiction explicitly with the sources on each side.

What we don't know

Each article includes three to five newsworthy unknowns: things the sources don't yet answer, follow-up questions, missing context. They're woven into the article rather than tucked into a footer. Uncertainty isn't an asterisk; it's part of the story.

How corrections work

Every Reed article has a Report inaccuracy link that opens the contact form. We read every report. When a reported claim doesn't hold up against the cited sources, we revise the article.

When new information emerges about an event we've already covered, the pipeline automatically re-runs the writing stage on that event. Articles aren't static records; they reflect the current state of the source pool.

We don't yet publish a public revision log per article. Until we do, external archives (Google's cached version, the Internet Archive) are the most reliable record of what an article said at a particular time.

Found something wrong in an article?

Report an issue

Who writes Reed

Reed articles do not carry individual bylines. Every article is produced by Reed's editorial pipeline — a chain of LLM-driven steps for source clustering, claim extraction, drafting, and review — running under human oversight at the system level rather than the article level.

We attribute every article to Reed Editorial. The attribution is collective by design: the integrity of an article comes from the verifiable sources behind it, not from the prestige of a name.

Editorial decisions — which categories Reed covers, how source types are weighted, what triggers a hold for human review, when the underlying AI model is replaced — are made by humans.

How the AI is supervised

The pipeline is staged so that no single LLM call decides whether to publish. Each article passes through six steps; failure at any one of them holds the article for human review rather than publishing it.

1
Triage
A code-level eligibility check against the source threshold above. Articles below threshold are dropped before any LLM runs.
2
Evidence
Sources are clustered by perspective (official, wire, established media, etc.) and de-duplicated by semantic similarity, so a story carried by ten outlets isn't counted ten times.
3
Research
For complex stories an automated web search runs to find missing context. New sources are added to the source pool with full extraction and entity resolution before any writing begins.
4
Claim extraction
Atomic claims are pulled from the sources with attribution. Contradictions between sources are detected and flagged; three to five newsworthy unknowns are identified.
5
Writing
The article is generated against the claim inventory, with mandatory confidence treatments per sentence — so every sentence ties back to a specific claim with a specific source.
6
Review
Code checks (claim coverage, format, length), a separate AI review that compares the draft against the claim inventory, a legal gate (Swedish patterns for harm, implied guilt, sensitive personal data, minors), a coherence check that scores the article 1–5 for internal consistency, and up to two revision passes.

The pipeline holds an article for human review when the legal gate flags it, when the coherence check fails twice, or when the AI reviewer escalates. Held articles wait in an internal review queue and don't appear on the site until cleared. For everything that does publish, human oversight sits at the system level — which categories Reed covers, how source types are weighted, which AI model the pipeline runs on — not on every article individually.

Mission and coverage

Reed exists to make verifiable original journalism cheap enough that it can be free at the point of access. We focus on Sweden and Swedish-language coverage with English translations for international readers. We cover what the source ecosystem covers — politics, business, culture, sport, science, and local news from Swedish municipalities — without paywalls, advertising, or a login.

For a longer take on why and how, see About Reed and Editorial standards.

Per-article transparency

Every Reed article has a transparency view that shows every source we used, every claim we extracted with its confidence treatment, all contradictions we found, and all the things the sources don't yet tell us.

Click transparency on any article. The transparency view lives at /news/[slug]/transparency for any published article.