This paper proposes SAGE, a statistical acceptance gate for self-evolving agents to filter out unreliable or regressive edits. It uses paired comparisons and one-sided paired tests to improve upon the naive rule-based gate.