The authors present SEA, an architecture that confines self-modification to a steering adapter and versioned harness around a frozen base model, admitting changes only through an anytime-valid gate that emits auditable certificates against a fixed error budget.
- Five loop controllers compose published guarantees, while verifier-in-the-loop mechanisms supply dense signal from issue text alone.
- On a 52-instance SWE-bench Verified subset, deliberate no-op-composite control isolates the suite's contribution at +4 for Glm 5.2 and +5 for Gpt.
- Event logs confirm that mechanisms fire and prevent regressions during single-run evaluations.
The system addresses the violation of learning-theoretic guarantees in self-evolving agents by ensuring modifications are auditable and bounded.