Check the news.
Check our work.
Software tests and a second AI reviewer are useful safeguards. Independent human review is how we measure errors.
Editorial qualification has not passed
No independent human accuracy report has been completed.
No independent human-reviewed qualification dataset has been completed. Automated tests check software behavior; they do not measure reporting accuracy.
Unreviewed claims are not counted as correct. Zero recorded invented claims with no reviews is no evidence of zero errors.
The results and the bar
| Check | Reviewed | Minimum sample | Observed result | Required result |
|---|---|---|---|---|
| Classification | 0 items | 200 unique items | Not measured news / Not measured category | 98% news / 95% category |
| Clustering | 0 pairs | 300 unique pairs | Not measured precision / Not measured recall | 95% precision / 90% recall |
| Brief accuracy | 0 briefs · 0 claims | 100 unique reviewed briefs | Not measured supported · 0 recorded invented claims | 98% supported / zero invented facts |
| Must-know recall | 0 references | 50 stories across 7 UTC dates | Not measured | 95% within 30 minutes |
How review works
A fixed, private queue records article references, clustering assignments and immutable brief versions before grading. A human reviewer checks every displayed factual claim against the originals, including any factual statement absent from the model’s claim list. Unsupported and invented claims lower the score. Review files retain judgments, reasons, source links and the queue fingerprint.
Sampling exposes quieter categories, rejected items, withdrawals and difficult same-topic pairs. These samples describe that cohort; they do not establish a population-wide accuracy rate or complete event coverage. Model configuration and per-record provenance must be checked together. The label importer requires human and independent-review attestations, but cannot authenticate the reviewer’s judgments.
What this report does not prove
- No measured accuracy superiority is established.
- Linked source evidence can be incomplete or mistaken.
- A passing sample would qualify that reviewed cohort, not guarantee every future claim.
Operations need separate delivered-alert, quota, billing-backstop and soak evidence. A passing accuracy report alone does not make the site ready for unattended launch.
Can readers help review?
Yes. Readers can report a specific claim and provide its original evidence using “Report factual problem” on a brief. These voluntary reports help identify errors, but an unreported brief is not necessarily accurate. Reports are self-selected, can be mistaken or coordinated, and do not measure an error rate. They never automatically approve, withdraw or train a brief. Independent, complete review of a fixed sample is still required for qualification.
Find a problem
Open a brief’s evidence inspector and compare its original reports. Review the public corrections and withdrawals. Our publication method describes source and claim checks.