Abstract
The more consequential a system becomes, the more carefully an institution should be able to answer a narrow question: what happens if a given AI provider changes its terms, its availability, or disappears entirely? This note outlines an architectural approach to model independence for institutional infrastructure.
A narrow but important question
Institutions adopting AI-assisted systems for consequential work are, whether they realize it or not, adopting a dependency. The relevant design question is not whether to depend on external model providers — most organizations will, at least in part — but whether that dependency is structured so it can be changed.
Designing for substitutability
An architecture built around model independence separates the institutional logic — the graph, the workflows, the decision structures, the audit trail — from any single model provider. Commercial models, private models, and local models become interchangeable components behind a stable interface, selected according to the sensitivity and requirements of a given workload.
AI as a component, not an authority
Model independence is one part of a broader principle: AI should function as a capability inside a governed institutional architecture, assisting judgment, rather than a black box that silently accumulates institutional authority. Human and institutional decision-makers retain the authority; the system retains the ability to change which intelligence provider sits underneath.
Status note
This entry is published as a working thesis. It reflects current thinking inside Solomonic Intelligence and is expected to change as it is tested against real institutional environments.