Reference model card

Purpose: a readable churn-risk prototype on fully synthetic no-contact data. Inputs: available purchases, support contacts and purchase recency over 30 days. Training: 300 independent customers. Feature means/scales fit on training only. L2-regularised logistic regression. Calibration/decision selection: separate 150-customer cohort. Sigmoid calibration and a threshold chosen with FP cost 2 / FN cost 10. Final evaluation: 150 later customers. Log loss 0.495936, Brier 0.169020, AUC 0.837216. Baseline: constant training prevalence, not final-test prevalence. Intended use: teaching and reproduction. Not for real customer contact, pricing or eligibility decisions. Known limits: the simulator resembles the model. One small cohort is shifted in time. No empirical fairness, censoring or deployment validation. Fallback: keep the baseline and defer when evidence is too weak. Risk does not measure treatment response. Owner: the learner running the prototype. A real release needs a named accountable operator.