You work on a two-sided payments network where payers send money to payees in real time. The team has built a new payment-routing policy that is expected to increase successful first-attempt payment completion by prioritizing rails with better recent acceptance and lower retry risk. Because payers and payees can interact repeatedly and liquidity conditions change over time, a standard user-level A/B test may create interference across sides of the network. You are considering a switchback test that alternates treatment by time block.
How would you design this switchback experiment so that you can credibly estimate the impact of the new routing policy while accounting for marketplace interference, time variation, and operational risk? Be explicit about the metric, power, randomization schedule, analysis plan, and the conditions under which you would ship or not ship.