A — 0: Serving is assumed to be easy after training. B — 1: The target environment is chosen, but load, latency and failure behaviour are untested. C — 2: Concurrency, input lengths, latency requirements, capacity, fallback and rollback tests are specified.
Evidence to attach: architecture and load-test plan. If weak: test a deployable baseline first. The inference engineering discussion is useful background for the questions, not proof of your requirements being met.