You already know: evaluation of LLM systems (golden sets, LLM judges), the LLM gateway, the guardrail gateway and the serving cases. This case designs the release process for LLM products: the system that decides whether a change ships. The design questions: what counts as a change, what to test before production, how large a regression a test can detect, how to trust automated judges, and how to roll out and roll back safely.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud; costs are illustrative.
1. The prompt, and the questions that change the design
A mid-level answer: "keep a golden set per product, run it on each release, ship if the score is above a threshold". A staff candidate asks what a release is (a prompt edit can break a product as badly as a new model, so every component must be versioned together), what size of regression a test can detect (a small golden set cannot see small drops), whether the judge can be trusted, and what happens after the offline gate (some failures only appear on real traffic).
Assumptions you state: