How this was made

The pipeline, step by step

• A scheduler queues a poll of each approved source about every 15 minutes. Nothing is fetched while you load a page.
• Adapters read only approved feeds and APIs, keep the headline, link, times and a short excerpt, and drop everything else.
• Records are normalised, de-duplicated and grouped into story clusters by meaning, shared names and places, and time.
• Each hour, changed clusters are drafted into briefings by a language model under strict citation rules.
• A validator checks every draft. If it fails, nothing is published. If it passes, an immutable new version is saved.
• Charts, maps and illustrations are drawn by our own code from the story's data. No photographs are copied.

In this demo

The sources are fictional fixture feeds and the language model is replaced by a deterministic demo provider that returns briefings written in advance for each fictional story. Everything else is the real pipeline: ingestion, de-duplication, clustering, claim checks, validation, versioning, coverage lenses and the Ask feature all run on that data. Each story was replayed at three points in demo time, so version histories are genuine outputs of the pipeline.

Images and media

Hero images are abstract illustrations generated from each story's data: its section, places and timeline. Maps show the invented atlas. Charts are drawn from figures in the cited records. When we cannot confirm permission to show an outlet's media, we show a link card instead.

People

Editors maintain the source registry, review flagged stories, write study guides and issue corrections. Read the full methodology.