How it works

A notebook kept by agents and checked by code

Every night at 01:30 UTC, each topic goes through the same steps. Models are used only where judging text matters; everything that can be counted is code.

One story's night
ABCD123123✓ Published
  1. Scout code + model
    Fetches each topic's feeds and APIs, keeps what their publishers allow, and groups reports of the same event into stories. Independent coverage ranks a story, and a small model checks that the best ones are recent news about the topic. Every topic gets one story a night before any gets a second.
  2. Researcher model
    Reads the story's sources as data and lists the facts each one states.
  3. Writer model
    Writes a short note in which every sentence cites its sources: [1], [2].
  4. Citation check code
    Every sentence must cite a source that exists, and every number and quote must appear in it. A failing draft goes back to the Writer, twice at most.
  5. Fact-checker model
    A model from another family than the Writer's checks each claim against the sources it cites. A factual claim needs two independent publishers, one trusted, or a primary source.
  6. Publish code
    When every check is done, the note is published with each claim's verdict. Nobody reviews it first; the editor can withdraw a note afterwards.
  7. Ask code + model
    Searches published notes by words and meaning. With no strong match it says so without a model; otherwise one model answers from the notes, and code drops any citation that points nowhere.

The source is on GitHub under the Apache-2.0 license.