Contextual bandit offer selection

Design a contextual-bandit system to choose, per user impression, which of many promotional offers to show, where a fixed A/B test wastes traffic on losing offers and offers churn weekly so you can never 'finish' an experiment. The reward (did the user convert) is delayed by hours-to-days. Walk through the explore/exploit setup, how you serve a policy at low latency, and how delayed and biased rewards complicate learning compared to a clean A/B test.