Plan the Evaluation of a Censored Time-to-Churn Model Before Launch

Your team has built a gradient-boosted Cox model and a discrete-time LightGBM model that predict churn timing for 200,000 subscribers. Roughly half the users in the data are censored. The VP asks for "the accuracy" and says the models will be used to (i) rank users for a retention campaign this quarter and (ii) estimate 12-month LTV for finance. A teammate reports "AUC 0.84" computed by labelling users "churned within 12 months" and dropping users with under 12 months of history.

1. What is wrong with that AUC, and what would you report instead for ranking? Name the estimators and how they handle censoring.
2. Which metric answers (ii), and how do you check it?
3. Design the validation split and the baselines.
4. The Cox model gets a concordance of 0.64 and the VP finds that unacceptable. How do you respond? What would you check before concluding the model is weak?