You’re given an anonymized DoorDash dataset at order-creation time and asked to predict late delivery risk (late = actual_dropoff_time > quoted_dropoff_time). 1) Define the target precisely and propose a time-based train/validation/test split that avoids leakage from future information (include exact cut dates). 2) Enumerate at least 10 high-signal, production-safe features available at order creation (e.g., store historical on-time rate by hour-of-week, dasher supply-demand index in zone, restaurant prep-time quantiles, distance and traffic, weather, surge/boost, customer lateness tolerance proxy). 3) Identify at least 5 leakage hazards and how you’ll eliminate them (e.g., features derived from post-pickup events, future average wait, features computed with non-causal windows). 4) Choose evaluation metrics for ranking and calibration (e.g., AUROC, AUPRC, Brier, calibration slope), justify thresholds for operational actions, and quantify business impact using a cost matrix. 5) Describe an online ramp: shadow mode -> treatment gating -> A/B test with guardrails; how you’ll monitor drift and recalibrate (e.g., Platt/Isotonic, periodic time-split retraining, population shift alerts). 6) Explain how you’ll handle cold-start restaurants/cities and seasonality (hierarchical pooling, entity embeddings, time-of-week effects).