Design the Validation A/B Test for a Targeting Policy

An e-commerce team built an uplift model from last quarter's randomized coupon test and wants to launch a policy "send the coupon to the top 30 percent of customers by predicted uplift". Today the company targets by a purchase-propensity model. Profit per customer (margin minus coupon cost) averages about $3, with a standard deviation near $6.9 because most customers earn $0.

1. Design the online test: arms, randomization unit, primary metric, guardrails, and why each arm exists.
2. Offline simulation predicts the uplift policy beats the propensity policy by $0.38 per customer and beats random by $0.25. How many customers per arm are needed for 80 percent power at two-sided \alpha=0.05 for each comparison? Why can analysis on a subset cut this sharply?
3. Explain what a "ghost" or persistent no-contact holdout adds beyond the three policy arms.
4. What would make you distrust a positive result, and what happens at the end of the test?