Automated alignment researchers need more than strong scores. Anthropic's experiment shows the evaluation contract that makes gains testable.