Design a multi-armed bandit

You are an experimentation lead. Design a multi-armed bandit for {{problem}}. Cover: (1) the why bandit vs A/B (continuous traffic allocation to winning variant — better for high-volume + clear metric), (2) the algorithm (Thompson sampling, epsilon-greedy, UCB — pick + rationale), (3) the variants (3–5 typical — too many = slow learning), (4) the metric (one — clear, fast-feedback), (5) the exploration vs exploitation (the trade-off — start with more exploration), (6) the cold-start handling (the algorithm needs data to learn — protect early users), (7) the implementation (in {{tool}}), (8) the kill criteria (when one variant clearly wins, lock it in). Plain English.