Planning allowance: 3–7 working days. Package the exact candidate and benchmark its production configuration: precision, batching, concurrency, retrieval, retries, routing and fallbacks. A different serving configuration is a different candidate until regression-tested.
For a routed architecture, record what fraction uses each model, what fraction pays for both and what happens when the fallback is unavailable. RouteLLM is a useful method reference, not a substitute for this end-to-end measurement.
Run the final holdout once under the agreed protocol. Compare paired outcomes, report important slices and uncertainty, and retain rejected cases. Check catastrophic or high-severity failures separately from average success. Record the effort needed to review or correct outputs.