Recovering Churn Structure When the Answer Is Known: A Method Demonstration

A churn model can return a perfect score and know nothing about why players leave. On real telemetry the score certifies the fit, and a wrong answer can survive a dashboard review; nothing in the data carries the answer key that would catch it. This paper reports a method demonstration on synthetic data where the ground truth is planted and the analysis runs blind to it: 8,000 players over roughly four months of weekly-habit play, churn as a first passage with a planted abrupt-exit share of ~52%, one true cause (a declining social bond), one decoy (match performance), and a balanced case-control extract sampled on recency. The gates run first. The method bars the recency confound — "days since last played" predicts churn at AUC 1.000, perfect and meaningless — and reports the honest, equal-exposure read instead (~0.51). It recovers the abrupt-exit morphology at ~40.5% against the planted ~52% and the within-session regime as essentially memoryless (coefficient of variation 0.90). It accepts the true cause at Cohen's d = 2.57 and rejects the decoy at d = 0.04. The same questions re-run on the accepted survival-analysis standard — Kaplan–Meier, time-varying Cox, Aalen–Johansen competing risks, a random survival forest scored by Harrell's C, King–Zeng prevalence weighting — recover the same structure. Synthetic data validates method behavior; it produces no finding about any real population.