Interleaved ranking experiments

Design an experimentation system to compare two search-ranking models quickly and cheaply, where a classic A/B test is too slow because the effect size per query is tiny and you'd need huge traffic to reach significance. The team ships ranker candidates several times a week and wants a fast, low-variance 'is B better than A' signal. Walk through how you'd use interleaving instead of (or before) a full A/B test, how you measure a winner, and how you guard against the failure modes that make interleaving lie.