WARNINGS ARE NOT VERDICTS: COUPLING-ROBUST ACTION CERTIFICATION FOR CITATION VERIFICATION

A citation verifier can correctly identify a missing detail yet wrongly downgrade a partially supported claim to insufficient evidence. We study this gap between detecting a defect and deciding to change a verdict. Our Projected-Interaction Action Filter (PIAF) tests each proposed change before it is applied. Holding the candidate fixed, PIAF requires its score-based advantage over the original verdict to remain nonnegative across all Fréchet-consistent couplings of local support and missingness and a declared family of losses. We characterize the largest acceptance set satisfying this condition and reduce the test to one-dimensional optimization and finite loss vertices. Negative advantages yield explicit counterexamples. On heldout GenSearch, PIAF retains 650 of 1,083 proposals and improves cluster accuracy by 5.15 points. A retrospective comparison at the same action budget finds one harm to a correct Partial verdict versus 35–51 for five confidence, calibration, and learned controls. A separate Learn-then-Test gate addresses population risk under exchangeability. On fresh GaRAGe, confirmation yields 7 harms among 74 selected actions; the one-sided 95% upper bound is 17.03%, below the declared 20% target at 7.64% question coverage. Separately, warning review reduces retained error by 7.72 points at 12.8% review on protected GenSearch. The results show why useful diagnostics need an explicit test of the changes they recommend.