Skip to main content

Fact checks

A fact check reads a document and checks the facts a reader would act on, such as prices, dates, addresses, counts, progress and who does what, against pages it finds and reads itself. It looks for the newest dated page from the party itself, so a figure that matches an old source can still come back as stale. It is advisory: it never publishes, changes or blocks the document. Agents ask for one with the review_artifact tool before research goes to someone outside the conversation. Through the API, submit the text to POST /v1/critic/reviews:
A fact check takes a few minutes, so REST and MCP answer at once with status: "running", a review_thread_id and a run_id. Read the run through the normal run API until it completes. The finished report’s status is completed, incomplete when the check stopped at its step limit before checking everything (its findings still stand), or failed when it returned no usable report. The completed report has one entry per lens. Each finding gives:
  • severity: low, medium, high or critical;
  • issue: the statement as written;
  • evidence: what the page shows, with its date;
  • source_url: the page the check read;
  • suggested_rewrite: the corrected text.
A finding whose source_url was not read during the check is marked unsupported: true; treat it as unconfirmed. The lens summary says what the check could not verify, and usage reports tokens, cost, searches and reads. From the CLI, submit the text and wait for the report: