AI agents write fluently whether or not they are right. Robby Prochnow decided that on his platform no AI assertion would ship unless it was tied to a source that could be checked live, and that the rule would fail closed: if the source cannot be confirmed, the claim does not go out.
He first enforced it in two places. Then he found that both could be routed around, so he added a third check at the very end, at delivery, that cancels anything still unsourced. Every agent on the platform is subject to it, with no exemptions.
By August 1, 2026 the gate had blocked 2,292 assertions. 1,819 of them were caught only at that final gate, after getting past the first two. That number is the argument for the third check: without it, those claims would have reached a person as if they were facts.