Evaluation Lab

Every release is judged against labelled cases, not impressions. Each run is compared with a baseline item by item, scores carry 95% confidence intervals because the test sets are small, and the judge model is itself checked against the labels.