Watch an AI get checked
This is a real run, not a mock-up. On 22 September 2026 an agent I built read 24 news articles and proposed labels for each one — what it is about, which industry, which country, which companies. Those labels would decide what a reader finds when they search.
A wrong label that nobody catches is worse than a slow correct one. So nothing the AI says is trusted on its own: a separate piece of ordinary code re-tests every single suggestion before it is allowed in. That checker is running on this page, in your browser, as you click.
Fixed rules run first
Before the AI is involved, plain code finds duplicate stories and matches company names against an approved directory. The AI can only pick from what that step hands it.
The AI suggests
It reads the article and proposes labels, each with a confidence score and the exact line it claims supports that label.
The checker re-tests
Does this label exist? Is that quote really in the article? Is the AI confident enough? Anything failing is dropped. No AI is involved in this step.
It goes somewhere
Filed automatically, held for a person, sent to a specialist, or dropped as a duplicate. If anything goes wrong with an article, it goes to a person.
Loading the run…
Where this fits
Why every build has a checker
The five safeguards this is one example of, the answer keys behind them, and a tuning change I reverted.
See the method The buildHow this agent is wired
Every step of the pipeline, what it cost per thousand articles, and the limits I would raise before anyone relied on it.
Open the build Open sourceRun this yourself
The two checker bodies running above are the same ones in the workflow. A replay on GitHub runs them over the recorded run and matches all 56 figures — no dependencies, no API key.
Open the replay