production-mlops · Governance & Production Readiness

Churn Prediction v2 — Governance Checklist

Model: churn-xgb-v2 · Skill: ml-governance-and-production-readiness · Command: /mlops-governance · Generated 2026-06-22
Risk Tier
MEDIUM
Model affects customer contact decisions. Misclassification can result in missed retention interventions for high-value customers. Human approval required before full promotion.

Risk Classification

Risk tierMedium
Decision typeCustomer retention targeting
Affected population~42k customers/day
Regulatory scopeGDPR Art. 22 (no auto-decision)
Approval requiredML Governance Board
Approval statusIn Review

Model Documentation

✓
Model Card complete
Architecture, training data, intended use, known limitations documented
✓
Data lineage documented
Feature pipeline → S3 → training split traceable via DVC
✓
Training reproducibility confirmed
Seed fixed, data snapshot pinned, artifact hash logged in MLflow
⚠
Intended use boundary — partial
Churn risk score defined; downstream CRM integration contract not yet documented

Fairness Audit

Segment Metric Value Threshold Status
Gender (M vs F) Recall parity 0.96 ≥ 0.90 Pass
Tenure cohort (<1yr vs >3yr) Precision parity 0.88 ≥ 0.85 Pass
Age cohort (<30 vs 30-60) FPR parity 0.14 ≤ 0.10 Partial — investigation ongoing
Geography (Urban vs Rural) Recall parity — ≥ 0.90 Not run — Sprint 14

Compliance Mapping

✓
GDPR Art. 22 — No fully automated decision-making
Model output is a risk score; final retention action triggered by human CRM agent. Compliant with prohibition on automated individual decisions.
✓
Data minimisation — feature set reviewed
17 features used; no special-category data (Art. 9) included. PII (email, name) excluded from training set.
⚠
Right to explanation — not yet implemented
GDPR Art. 22(3) requires meaningful information on logic when model influences decisions. SHAP explanations planned for Sprint 15.