The paper, arXiv:2603.23749v1, examined evaluation efficiency across eight benchmarks, 33 agent scaffolds, and more than 70 model configurations. The authors found that standard full-benchmark evaluation is expensive because each run requires interactive rollouts involving tool use and multi-step reasoning. A key finding was that while absolute score prediction degrades under what the authors term 'scaffold-driven distribution shift'—where performance varies depending on the framework wrapping the underlying model—rank-order prediction remains stable across shifts. Exploiting this asymmetry, the researchers proposed an optimization-free filtering protocol: restrict evaluation to tasks with historical pass rates between 30% and 70%. This 'mid-range difficulty filter,' grounded in Item Response Theory, reduces required evaluation tasks by 44–70% while maintaining high rank fidelity under both scaffold and temporal distribution shifts. The protocol outperformed random task sampling, which the authors noted exhibits high variance across seeds, and also outperformed greedy task selection under distribution shift conditions. The authors concluded that 'reliable leaderboard ranking does not require full-benchmark evaluation,' suggesting significant cost savings are achievable for AI agent leaderboards without sacrificing the integrity of model comparisons. The primary source is available at https://arxiv.org/abs/2603.23749.