Can AI systems be allowed to improve themselves if they are also allowed to judge their own improvement?

SFA-Bench White Paper v1.0: A Grammar for Governed Improvement

This paper introduces SFA-Bench as a governance benchmark for controlled AI/system improvement. It focuses on how proposed improvements should be evaluated when the system proposing the change must not also control the judge, the evaluation conditions, the ratification process, or the lineage record.

The release includes the governed improvement path: frozen-zone protection, pre-registration, controller execution, external candidate harnessing, human ratification, lineage and rollback, circuit breakers, adversarial candidate testing, and a reviewer-facing research release pack.

The central claim is bounded: SFA-Bench does not claim autonomous self-improving AI, proof of alignment, or general AI safety. It proposes a reproducible governance grammar for deciding when improvement claims are legitimate, auditable, replayable, and safe to promote.

The question this work raises for the community is:

Should future AI governance frameworks require protected evaluation, human ratification, lineage records, and circuit breakers before AI-generated improvements are accepted as evidence of progress?