Agree that some of the claims are forward-looking. The messiness of the real-world and real-user fact checks. No ground-truth verdicts are provided or used in the study though. It only measures the level of agreement between the selected models, not which one is right on which claim. I.e. none of the claims is actually labelled.