“High-quality, uncontaminated evaluation datasets are often neglected but are as important as training data for rigorously measuring and advancing the capabilities of general AI agents.”