The paper outlines fixed rules meant to constrain automated research when machines may act faster than institutions can respond.
The proposed design has two control domains. A Human Authority Boundary preserves human responsibility, discretion and institutional approval, while a Machine-Enforceable Control Boundary restricts protected changes using explicit authorization conditions, independent checks, records of evidence, limited commit points, revocation and recovery controls.
The paper also describes a reference control path for automated AI research and development, possible attacks and failures, evaluation protocols, deployment scenarios and implications for standards. It argues that watching a system, collecting evidence or verifying an action should not by itself authorize the action.
The proposed control rules
The researchers propose a deterministic dual-boundary architecture for automated AI research and development. Its machine-enforced boundary is designed to control protected system changes through authorization rules, independent verification, preserved evidence trails, scoped commit boundaries, revocation, localized failure handling and fresh checks before readmission after recovery.
The architecture keeps human acceptance and institutional responsibility separate from technical execution. The paper also presents control invariants—rules intended to remain true during operation—and a model for securing AI systems that monitor other AI systems. The abstract does not report experimental results or a measured improvement over existing systems.
Why fixed limits matter
If AI-assisted research cycles become faster than ordinary human and institutional response, observation and reporting may not be sufficient as the main real-time safeguards. This proposal addresses that gap by placing machine-enforced limits around certain system changes while retaining human authority over acceptance and responsibility.
Its central relevance is organizational as well as technical: it distinguishes seeing an action from authorizing it, verifying it from carrying it out, and recovering a system from restoring its previous authority. Those distinctions could inform discussions about controls and standards for highly automated AI research.
Evidence and caveats
This is a research-level proposal published as an article in the Zenodo repository, not a report of a completed deployment or experiment in the abstract provided. The paper describes control rules, failure modes, evaluation plans and deployment scenarios, but the abstract gives no test results, performance measurements or evidence that the architecture works in production.
The publication intentionally omits unpublished patent claim language, confidential claim charts, private source locations, provider-specific production settings and other non-public implementation details. Its practical effectiveness therefore remains to be established through the evaluation protocols it proposes and through real-world testing.
// Source
Zenodo (CERN European Organization for Nuclear Research) · 2026 · DOI: 10.5281/zenodo.23031785
Authors: The Second Waters
