The editorial lens

SPHRE writes for a reader anywhere. The working test for every page: a reader in Tehran, Moscow, Beijing, or Caracas should see data and history, not someone else’s narration of their country. A page that fails that test is wrong even when every number on it is accurate. Rule by rule, the policy states whether it is enforced by a check in the build or held as an editorial commitment.

Relevance is judged from the country’s perspective

A country’s timeline and dossier answer what matters to that country and the people who care about it — never where it appears in a great power’s story. Which countries a historical event belongs to is a curated editorial decision made under that rule: who lived it, not who read about it.

Whose voices carry contested events

The event’s own protagonists speak first, and world-scale events carry the dominant perspectives side by side — no country is the default narrator. The sourcing law is mechanical: a quote is published only when its passage appears verbatim in a fetched primary source. The balance itself is an editorial judgment, reviewed against each event’s speaker roster. When no verifiable archive carries the other side’s voice, the event ships what verifies and records the gap.

Estimates keep their names

Where more than one institution estimates the same quantity, each estimate renders as its own series — layered, never blended, stitched, or averaged — and the provenance line names each estimator. That plurality is enforced by a build check; the prose rules — estimation verbs for modeled series, provisional language on the newest figures — are commitments held by review, not yet by a script.

The verification ceiling

Machine-drafted work is grounded on the first and latest observations of each series it uses, read from audited source snapshots, and the pipeline is built not to write past that ground: endpoint claims are supportable, path claims — acceleration, records, reversals — are treated as unsupported, and every numeral must trace to those observations or to the dated news peg the piece names. A claim the pipeline cannot ground does not publish; a draft that still is not clean after one revision is removed and logged publicly. The gates are documented on the methodology page.

One voice, human or machine

SPHRE has one register regardless of byline, and machine-drafted pieces are labeled where they appear. The draft is reviewed by a second model required to be a different model id than the drafter — both ids currently come from the same model family, served through the same provider — and the published check record names both ids.

Where coverage is uneven

Brief topics rotate through a fixed, code-defined roster of themes matched against an open news index, so coverage follows data availability — and data availability is not even across the world. Where the data ends, the policy is to show the gap rather than paper over it. See About for the publication and Sources for who produces and funds the data.