AI Network Evolution Assurance: From Capability Roadmaps to Dependency Roadmaps

AI-native networks are planned on roadmaps. Operators and their suppliers state, several years in advance, how integration will deepen, how much of the network will run under autonomous control, and what performance will follow. An industrial adopter deciding whether to build a long-lived system on that network can evaluate those prospective capabilities in advance. The same evolution does something else that the adopter cannot read. Each generation of joint optimization, revalidation, and accumulated operational state changes how easily that adopter could later move to a different implementation. No standardized measurement discipline or routinely maintained forward projection comparable to the capability roadmap exists for it, and for an adopter whose assets outlive several network generations, the resulting uncertainty can be enough to hold back integration that would otherwise be worth having. This paper treats that gap as an engineering problem and proposes AI Network Evolution Assurance. We distinguish two mechanisms that are usually treated together. Evolution Observation and Assessment measures the dependency the network is actually forming and maintains a forward envelope for it; Evolution Intervention changes the dependency path itself, through portability work, alternative-implementation validation, interface redesign, and similar measures. Observation and no assurance at all share the same underlying dependency path and differ only in what is known about it. Intervention produces a different path. A network-specific case separates dependency created by the operator's own internal evolution, by explicit cross-domain integration, by propagation from the former into the latter, and by implementation practice that is comparatively easy to undo. Numerical experiments then ask how deep integration can go as the range of industries the architecture must support widens, which kind of dependency makes intervention worth its cost, how far assurance evidence can be shared, and what happens when the operator advances its own architecture without the adopter choosing anything. In the illustrative baseline the adopters differ in one respect only — how they evaluate dependency — and the sustainable depth without assurance falls to zero once safety-certified operations are in scope. Observation recovers most of that range; intervention extends it to the edge of the range considered. Observation is not, however, arranged to favour integration. Past an actor-specific severity threshold — which reduces to that actor's own tail multiplier in the unclipped linear case — its advantage disappears and, where there was a positive no-assurance baseline, reverses. Under imperfect observation the two mechanisms separate cleanly: the incremental value of observation falls to zero as information quality vanishes, while intervention retains substantial value because it changes what happens rather than only what is known.