Models are replaceable. Mandates persist.

Every architecture makes a decision about where its memory lives. When the only useful record sits inside a provider-specific conversation, replacing the provider means reconstructing not only the execution environment but also the organization’s understanding of its own work.

A provider-neutral operating layer takes a different position. The mandate, authority, evidence and decision history remain above the individual runtime. A model supplies a capability; it does not own the reason that capability was invoked.

Neutrality does not mean pretending every provider is equivalent. The right system should be opinionated about eligibility, current evidence, failure cases and the tasks for which a capability has actually been evaluated.

A registry is therefore more than a list of models. It is a record of scoped decisions: which capability is eligible for which role, based on what evaluation, at which revision, with which conditions for rollback.

The practical test is simple. Could another actor understand the mandate and continue the work without inheriting a private, undocumented conversation? If not, the organization has rented more than intelligence. It has rented part of its own operating memory.