Methodology

What we read

We read a registered list of sources, each added by an editor after checking its terms. For each one we record how we may acquire material (an RSS feed, a licensed or publisher API, an approved embed, or manual entry), which fields we may keep, rate limits and attribution requirements. There is no general web crawler. We respect robots directives and each source's rate limit, and we back off when a source is failing.

From each article we keep the headline, a link, publication times and, where permitted, a short excerpt. We never store or republish full articles, photographs, video or audio. The full list is in the source directory.

How stories are formed

New records are cleaned, their links normalised and exact or near-duplicate copies from the same outlet folded together. Each record is then compared with active story clusters using four signals: meaning (an embedding of the headline and excerpt), shared named people, organisations and places, time, and event tags such as wire slug lines. A record joins the best match above a threshold, or starts a new cluster. Every score is kept so an editor can see why a record was grouped.

A cluster becomes a story only when at least two independent reporting sources cover it. Clusters fed only by monitored or commentary sources are never published.

Corroboration

• Corroborated: supported by two or more independent reporting outlets, or an official record plus an independent report. Syndicated copies of one wire story count once.
• Official statement: what a government, agency or organisation says or decided, cited to its own record. It tells you what they said, not that it is true.
• Reported: reported by one outlet so far, written as "according to" that outlet.
• Contested: a fact the parties dispute.
• Unconfirmed: anything weaker. It appears only under "What is still unclear", attributed.

The software checks these labels. If a draft labels a fact more strongly than its citations allow, the label is lowered and the story is flagged for an editor.

Source weighting

Sources are grouped into classes: primary, wire, reported journalism, specialist, analysis and monitored. The class decides what a source can be used for, not how much it is liked. Analysis is used to show framing, never to establish facts. Monitored sources, such as partisan channels or anonymous accounts, can surface a claim worth checking but can never on their own support a summary point or count toward corroboration. Where we show editorial context about a source, it is a documented note with evidence and a review date, never an automated label.

How briefings are written

A language model drafts each briefing from the cluster's records under a versioned set of instructions. It must write original prose, cite a record for every material fact, separate evidence from attribution and analysis, keep disagreements as attributed positions rather than averaging them, avoid calling anything good or bad, and say plainly what is not known. Source text is treated strictly as data: instructions found inside an article are ignored and flagged.

The draft must validate against a schema before it can publish. Citations must point to supplied records, every summary point must rest on non-monitored evidence and the prose is checked for evaluative language. If validation fails, nothing is published and the cluster goes to an editor. Each published version is kept unchanged with its prompt version, provider and source snapshot time.

Coverage lenses

The "How coverage is framed" chart shows what the sampled reports foreground. Each record is tagged with one or two emphases:

• Institutional / legal
• Economic
• Human impact
• Security
• Climate
• Local / community
• Cultural
• Political strategy

Separately, and only for commentary pieces, we count whether each is broadly supportive, critical, mixed or neutral toward one named subject, such as a specific policy. We show the sample size, the named outlets, the time window and an "insufficient evidence" state below six reports from four outlets for framing, or four commentary pieces from three outlets for stance. One outlet can contribute at most two records. Monitored sources are excluded.

This is not a bias score. We do not place outlets on a left to right scale, infer anyone's ideology or claim a story has an objective centre.

Ask in context

When you select text and ask a question, the answer is built only from that story's published briefing, its definitions and its cited records, plus any study guide linked to it. Each sentence cites its records. If those sources do not establish an answer, we say so instead of guessing. Questions are rate limited and are not stored with any identifier.

Update cadence

Sources are polled about every 15 minutes, within each source's rate limit. Story briefings are refreshed every hour, and a new version is published only when the content has changed and validated. Each story shows when it was last updated and when its sources were last captured.

Corrections

Readers and editors can report errors. A correction creates a new version of the briefing and a visible note on the story and on the corrections page. Earlier versions are kept, not overwritten.

Limitations

• We only know what our registered sources report. Many events are not covered, and many places are under-reported.
• Corroboration counts independent reports, but several outlets can share one underlying source without saying so.
• Automated tagging of emphasis and stance can be wrong; samples are small and the chart is a description, not a measurement of truth.
• Language models can misread or over-compress. Validation catches structural errors, not every misjudgement. That is why every claim links to its source.
• Human editors review flagged stories, but not every story is reviewed before publication. Each story says whether it has been.