How would you promote managed AI configuration across environments?

I distinguish resources that can be recreated from stateful assets such as indexes, evaluation evidence, fine-tuned models, or provider-managed conversations that need migration plans. Quota and marketplace access may require pre-provisioning outside the deployment transaction, so readiness checks fail early. Rollback accounts for model retirement and irreversible data changes rather than assuming the previous endpoint still exists. The release manifest makes a production response attributable to both application code and managed-platform state.