Boundary-Preserving Alignment: From Bounded Artifacts to Bounded Authority: Implementing Safety in Long-Horizon Human-AI Systems Through the Governance of Authority Promotion

This position paper asks when an AI-generated artifact may be used as something more consequential than the output it began as. A summary can be useful without being a quotation; an approved plan is not necessarily authorized for execution after conditions change; and a successful tool call is not proof that the intended outcome occurred. The paper proposes boundary-preserving alignment as a complementary layer of AI safety. It follows artifacts from generation through display, execution, verification, persistence, publication, and reuse. At each transition, the system revalidates provenance, scope, uncertainty, permitted use, and authority. Human approval is treated as a finite-bandwidth control input, not as a guarantee of meaningful oversight. The framework distinguishes false promotion of unjustified artifacts from false suppression of legitimate ones. It offers design recommendations, evaluation hypotheses, and fault-injection cases for long-horizon human-AI systems. It does not claim to solve model alignment or to demonstrate the effectiveness of a deployed system.