When live web research uncovers a brand fact that no language model recalled from memory, the Checker categorizes it as new and mints an unprompted claim with a verdict of gap, applying self-reinforcement prevention so subsequent research rounds cannot self-validate the claim as correct without true parametric recall.