How our AI desk works
The desk follows the steps below before publishing. These checks reduce unsupported claims, but they cannot guarantee that every source or model response is correct.
Ingest and grouping
Feeds are read on a schedule that runs every minute. The reader keeps the headline, the link, the publication time and the article text needed for the language model. The text is used for checking claims and is never reprinted in full on the site.
Reports about the same event are grouped, with a 48 hour window so follow-up coverage joins the original story instead of starting a new one.
The source gate
A group needs three publisher websites, or two publishers with a separate primary source, including a publisher in the higher trust tiers. The gate groups related website domains so the same site does not count twice. Groups that fail the gate can still appear on relevant timelines, without a story page.
Facts, data and writing
A first model call extracts candidate claims from the source text. A second call checks each candidate against that text and discards the ones it cannot confirm. Confirmed claims are stored with the sentence that supports them.
A separate step adds data the desk already holds: the price move of the tagged coins, trading volume, protocol totals from DefiLlama, related unlocks and events, and prior incidents. The writer receives only confirmed claims and stored data. It has to return structured sections, a short answer, a frequently asked questions list and metadata, never page markup.
Validation and publishing
Before publication the story is checked again: numbers against stored values, citations in every paragraph, length, banned wording, and similarity against source articles and our own recent stories. A rejected draft goes back to the next provider with the reasons attached. If no provider produces a valid story, the draft is logged and nothing is published.
Publishing has daily and hourly caps and a hold switch. The operator can pause the desk, hold a story or unpublish a page without editing code.
Corrections
Successful rewrites add a dated update or correction to the same story address. The log identifies the new evidence. A failed rewrite leaves the last published version in place.