Routing and escalation direct each request to the cheapest model able to handle it, then move to a more powerful model if the result fails validation or if the task requires it. Local is preferred and the cloud is used last.
Example
Local model first; if validation fails, retry with a cloud model.