Governed Metaprogramming: Reclassifying Eval as a Controlled Effect for AI Systems

A recent paper on arXiv (2605.05248) introduces the concept of governed metaprogramming, which redefines the eval operation from an unrestricted primitive to a governed effect within intelligent systems. As AI technologies increasingly create executable code in real-time—such as large language models generating programs, agents developing workflows, and systems enhancing their own functions—the shift from code representation to execution amplifies authority. The authors contend that this shift should be regulated like any other effect. In their framework, program representations (machine forms) are treated as first-class values, manipulation of forms is considered pure computation, and the process of materialization (transitioning from form to executable machine) is a governed effect that undergoes structural scrutiny. The governance system evaluates the program's capability needs, adherence to policies, and resource assessments prior to allowing execution.