About

Every news event has multiple simultaneous coordinates. The same earnings call is a Technology sector story (GICS), a corporate action (NOSIBLE Corporate Events), an Analysis piece (IPTC Genre), an economic-frame article (Media Frames), and — depending on the quarter — a Joy or Surprise in tone (Ekman). A single taxonomy captures exactly one of those dimensions and ignores the rest.

Legacy news taxonomies like IPTC Media Topics were designed for desks, not for analysis. They solve the routing problem (get this article to the right editor) without solving the research problem (find every event with a particular combination of sector, geography, political action type, and emotional register across the last three years). NOSIBLE World was built to solve the research problem.

The platform classifies every event across 13 complementary ontologies simultaneously — topical, sectoral, political, cyber, health, sports, cultural, framing, emotional, provenance, and more. The full list and reasoning is in the Methodology. Each ontology answers a different question. Together they give analysts the ability to pivot between coordinates — from country to sector to political action type — and surface the events that no single-axis system would show.

The classifier is not a monolithic LLM doing everything. It is a multilingual OpenAI embedding classifier - one text-embedding-3-large vector over the event title and description, scored by cosine similarity against taxonomy caches. The top-ranked label per taxonomy ships. There is no LLM in the classification path.

NOSIBLE's search engine continuously ingests millions of articles from the open web across every major language and writes pre-built daily snapshots. NOSIBLE World consumes those snapshots through a nine-stage pipeline: ingest, cluster with Leiden community detection, same-incident deduplication, entity-aware splitting via spaCy NER, cross-lingual merging, cross-day story chaining, single-call LLM enrichment (one Gemini call per event, not per article), and multi-ontology classification via the embedding ensemble described above.

Events carry a deterministic identifier derived from their date, IAB category, geography, and topic code — reproducible from the event record alone, collision-resistant across 20+ years of events, stable under re-runs of the same day's data. Every artefact is written atomically via tmp-then-rename so the front end never reads a partial file.

The data model is public — the full specification is committed to this repository as DATA_MODEL.md. No proprietary encoding, no undocumented fields. Analysts who want to query the raw data programmatically can use the enterprise API or contact us about a direct feed.