SAGE Paired Test Cuts Self-Evolving Agent Regressions from 36.5% to 0% on LiveMath

The arXiv paper defines SAGE as a conservative filter on top of existing optimizers. It runs identical validation items through both current and edited skills, records per-item outcomes, then applies a one-sided paired test before acceptance. This directly addresses two documented failure modes: permanent regressions on previously solved items and upward bias from the Optimizer's Curse on finite noisy sets. Across five benchmarks and four LLMs under equal-budget conditions, SAGE lowered regression rates in 19 of 20 configurations. LiveMath regression fell from 36.5% to 0% and OfficeQA from 42.8% to 0% with DeepSeek-V4; final scores rose in every setting, reaching 48.78 on LiveMath versus the prior 34.15. The method recovers baseline behavior exactly at its boundary parameter. Magnitude (YC S25) inference engines that rely on persistent skill documents will adopt equivalent statistical gates to stabilize autonomous loops. Without them, self-edits remain vulnerable to silent breakage even when average metrics improve. Deployment records from prior agent frameworks show that unchecked regression compounds across iterations, eroding reliability faster than raw capability gains can compensate. Operational adoption requires only a validation harness and a fixed significance threshold; no additional training is needed. The gate filters roughly half the edits a naive scorer would accept while preserving or improving end performance.