Diagnose CTR drop after recommendation launch

A horizontal recommendations carousel recently went live on the home page. In the post-launch experiment, the Treatment group exhibited lower home-page click-through rate (the primary interaction metric), while app-wide daily active users and total time spent stayed flat.

Tasks:

1. Specify success metrics and guardrails for this surface: primary metrics should be qualified CTR or saves per impression, and guardrails should cover session length, bounce rate, latency, crashes, and notifications sent. Include variants normalized by exposure and by eligibility.
2. Lay out a step-by-step diagnosis plan covering instrumentation validation (event drops, duplicate fires), exposure parity, novelty and position bias, cannibalization of other entry points, how often the carousel is shown, content quality, scroll-depth and viewport effects, cold-start personalization, ranking changes, and infrastructure incidents. Indicate the exact logs or queries you would run.
3. Suggest at least eight segmentation cuts to localize the effect, such as new versus returning users, geography, device, app version, network quality, time of day, content domain affinity, session depth, notification-referred versus organic, and paid versus organic users.
4. Recommend next steps: targeted fixes or follow-up experiments (for example, cap surfacing frequency, change the default slot, diversify content, or boost cold-start personalization), and explain how to decide between rollback and iteration using pre-registered thresholds.

Overview: This question tests a data scientist's ability in experiment diagnosis, defining metrics and guardrails, validating instrumentation, performing segmentation analysis, and applying causal reasoning to recommender-system effects.