Verify the functional correctness of an LLM-based recommendation system

Build a labeling rubric with 4-5 dimensions (relevance, freshness, novelty, diversity, serendipity). Run offline eval on a held-out slice of historical traffic. Compare to online metrics via a sandbox / interleaving experiment. Define guardrails against regressions (latency, repeat rate, click-through to wide variety). When a new model ships, run the offline eval and a 5% shadow. Promote only if metrics hold and shadow traffic has no quality regression.