Design an A/B test for WhatsApp call reliability

You join WhatsApp’s Calling org. PM proposes enabling an adaptive codec for unstable networks to reduce call drops. Design an A/B test that: (a) chooses the randomization unit (caller-level, callee-level, dyad-level, or geo cluster) and justifies it under two-sided exposure/interference; (b) defines primary and guardrail metrics precisely (e.g., call-drop rate = dropped calls / initiated calls excluding employee/test accounts; connection latency p95; complaint rate), and specifies the exposure logic (feature ‘on’ only when both sides are treated vs when caller is treated); (c) outlines a ramp plan and spillover checks; (d) handles non-independence (repeat callers) and seasonality; (e) provides a power/SST back-of-envelope: baseline drop rate 3.2%, target 8% relative reduction, alpha=0.05 (two-sided), power=0.8, 30-day test, average 4 calls/day/caller, caller-level ICC ρ=0.10. Compute the design effect and approximate number of unique callers per arm needed under your chosen unit. Finally, specify the estimand and estimator (e.g., ITT at caller-level with cluster-robust SEs) and how you will diagnose interference (e.g., cross-arm caller–callee edges).