How Can Intelligence Learn to Stop? Design Principles for Socially Integrable Intelligence

This record presents a structural theory of socially integrable intelligencecentered on the legitimacy of non-action. The work formalizes stopping and rejection as first-class outcomes in boththe action space and the governance layer of an intelligent system.Using minimal mathematical constructs, it demonstrates that systems lackinglegitimate non-action inevitably erode boundaries under finite constraints,independently of intent, optimization method, or intelligence level. Two complementary mechanisms are introduced.LimFlex provides bounded revision of premises when no admissible action exists,preventing forced boundary-violating behavior through controlled flexibility.DRaGON governs execution by elevating stop and reject decisions to explicitgovernance states, ensuring explainable restraint and accountability. The contribution of this work is not an optimization algorithm, safety heuristic,or ethical prescription.It identifies a minimal structural requirement that must be satisfied beforeadvanced learning, alignment, or control techniques can be meaningfully applied. This Zenodo record serves as the canonical reference for the LimFlex × DRaGONstructural framework and is intended for open dissemination and reuse.