AI-model optionality that needs re-verification isn't optionality

A control plane solves the AI-model routing problem. It does not solve the authority problem.

Route between providers. Abstract the API. Add a gateway. The logic is clean: if the dependency is on a specific model, remove the specificity. This is necessary work.

Now try the next step. Route a prompt to a cheaper model. The call succeeds. The output is structurally valid.

Can that output be promoted under the same policy as the original? Can you demonstrate equivalence without a manual review?

If the answer requires a human to reconstruct it, the control plane left the structural dependency untouched. You solved routing. Authority is still implicit inside the value stream.

Optionality that depends on someone re-verifying every substitution is not optionality.

It is operational cost wearing an abstraction layer.