A product has **users with multiple accounts**. In the UI, these accounts are shown as a list. - **Current ranking**: accounts are sorted by **most recent visit** (most recently visited account appears at the top). - **Proposed change**: sort accounts by **number of notifications** (accounts with the most notifications appear at the top). You are asked: 1) **Historical/offline analysis**: Using existing logs and historical data, how would you assess whether this ranking change is likely to improve the product? Be explicit about: - what success means (primary metric + diagnostic metrics + guardrails) - what data you need (events/logging schema at a high level) - key confounders/biases (e.g., selection/position bias, heavy-user effects, regression to the mean) - what analyses you would run and how you would interpret results 2) **A/B test design**: Propose an online experiment to measure the causal impact of the new ranking. Cover: - experiment unit and randomization level (user/account/device) - eligibility criteria (e.g., users with 2+ accounts) - key metrics (primary/secondary/guardrails) and how to compute them - sample size / MDE considerations and duration - analysis plan (e.g., CUPED, handling multiple comparisons) - pitfalls (spillover, novelty effects, instrumentation, SRM) and how to mitigate them