policyfactneutralRetrospective supervisory oversight becomes ineffective as a risk management tool for autonomous AIArtificial Intelligence28 Jul 2026http://arxiv.org/abs/2607.21345v1