“Current model leaderboards like SweeBench are insufficient for evaluating real-world performance due to test set leakage and reliance on subjective human preferences.”