Beyond Scaling III: Runtime Topology and Governed Self-Growth — A Transition-Oriented View of Governable Intelligence
Year: 2026
Type: preprint
Abstract: As intelligent systems become more capable, their risk profile is no longer captured by capability growth alone. The more important question is how runtime conditions change as systems act, delegate, recover, revise, and reshape the possibilities available to future action. This paper extends the structural transition view introduced in Beyond Scaling II and frames runtime topology as a post-scaling governance object. In this view, an intelligent system is not only a model that produces outputs. It is a participant in a moving field of dependencies, evidence, recoverability, responsibility, and future access. The paper argues that governance after scaling must preserve proof-level dependencies across runtime transitions. A transition may remain locally acceptable while weakening the conditions that make future action observable, recoverable, or revisable. Four concepts form the core of the paper: 1. Runtime Topology — runtime is treated as a moving structure of dependencies, evidence, recoverability, responsibility, authority, and future access rather than as a flat execution layer. 2. Proof-Level Dependency Preservation — claims such as “authorized,” “recoverable,” or “bounded” depend on relations that must remain intact across runtime change. 3. Runtime Boundary Formation — acceptance under one runtime configuration does not automatically remain valid after the conditions that justified it have changed. 4. Governed Self-Growth — learning and adaptation remain governable only when reflection, evidence, constraint, counterfactual review, revision, and recovery remain separated from direct self-relaxation or direct self-rewrite. The paper does not present a deployable governance architecture, operational boundary-generation algorithm, internal coordinate system, or universal safety proof. Its contribution is a transition-oriented framework for reasoning about runtime governability when intelligence begins to alter the conditions of its own future action.
Author: Suzuki, Noriyuki
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7204196574
Cites:
Cited by:
Related to: