Epoch

Agent Leaderboards Keep Skewing: High Scores Do Not Prove Capability Gaps

Agent benchmarks face flawed tasks, leaked answers, and sampling error. Judge rankings by confidence intervals, significance, and sample size—not cham