Propose an online experimentation plan to evaluate the file recommender in production across multiple organizations where collaborators can influence one another. Specify: primary metrics (e.g., file open-through-rate, time-to-open), secondary/business metrics (productivity proxies), and guardrails (latency, error rate, privacy incidents, access denials). Choose the unit of randomization (org-, team-, or user-level) and justify to minimize spillover; describe bucketing, stickiness, and holdouts. Compute required sample size and MDE with clustering/ICC assumptions; select variance reduction (CUPED/stratification) and sequential monitoring approach with alpha spending. Detail ramp schedule, novelty and carryover controls, and interference detection. Define logging needed to reconstruct exposure and attribution, plus a difference-in-differences fallback if only partial randomization is possible. Explain stop/ship criteria and how to guard against Simpson’s paradox across tenants and roles.