The evaluation gate can verify that protected files changed, that the report matches the exact diff, and that each stage has quality results and a qualifying sample size, but it cannot verify that the evaluation actually ran, that the numbers are genuine, that the thresholds are reasonable, or that the golden dataset covers genuinely important failure modes, so each report still requires human review.