Methodology

Sources

The briefing reads from a registry of news outlets curated for country breadth, language coverage, and editorial spectrum. Every source is labelled with ownership type (state, billionaire, private, public, cooperative, nonprofit), bias lean, and a reliability rating. See the source list.

State-owned outlets are labelled, not excluded. Both direct publisher (e.g. Dow Jones & Company) and top-of-chain owner (e.g. News Corp / Murdoch Family Trust) are surfaced.

Translation

Foreign-language sources have headlines and ledes translated at ingest, with the original-language text preserved alongside. Cluster canonicals get full-body translation. All translation runs locally on the publishing machine; no third-party API is ever called.

Claim extraction and attribution

From each canonical article, the pipeline extracts atomic factual claims — paraphrased ≤30 words, anchored to a literal quote ≤25 words. Every claim is attributed inline to a named source. We never produce bare assertions.

When multiple outlets in different ownership categories cover the same event (clustered at ≥80% headline+lede similarity), the claim ledger lists all of them as corroborating sources. When only one source reports an item, that's surfaced as "single-source story — corroboration pending."

AI disclosure

The synthesised prose — story headlines, deks, body paragraphs, perspective panels, claim callouts — is generated by an AI editorial pass under written rules (third person; attribution first; no invented facts; quotes capped). Source attributions and bias-lean labels are not AI-generated; they're structured data from the source registry.

Per EU AI Act Article 50 and California AB 3030, this page marks the AI-generated content explicitly. Disputed claims go through corrections.