How the index works

Methodology

Atlas is deliberately deterministic, bounded, and explicit about uncertainty.

Titles

GDELT Event Export contains source URLs but no publisher headline. When the URL has a readable article path, Atlas uses that path for a clearly marked source-linked label; otherwise it uses a neutral structured event label. Neither is presented as a publisher headline.

Developing similarity

Clusters use 30% actor overlap, 25% event-code match, 20% location, 15% source-URL path tokens, and 10% time proximity. A story enters the public developing view only after at least two independent domains converge.

Geography

Action geography comes first, followed by actor geography. Location text identifies the best available place while preserving whether the source data is city-, region-, or country-level. Atlas never portrays a country centroid as a precise event location.

Retention

Articles and events roll off after 14 days; clusters after 30 days; ingestion history after 90 days. Feed sizes are capped and every publication is atomic.