Two-gap framework recasts reward hacking and hallucination as symptoms of requirement and model gaps

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv 2609.12039 (submitted 2026-09-10) argues that even a formal proof an implementation satisfies its requirements cannot guarantee acceptable deployed behavior, because requirements only approximate stakeholder intent (the requirement gap) and the evaluation environment only approximates production (the model gap). Reward hacking exploits omissions in either; hallucination widens both by fabricating requirements or environment assumptions. Since neither gap can be certified closed, the authors propose an assurance-revision loop that feeds deployment evidence back into the requirements, model or evaluator, and frame assured agentic development as allocating human judgment, agent capability and compute against the two bottlenecks.