Predictive human-risk ML model

[jusso-dev]

Why

Today's risk score is a fixed formula. Competitors use trained models on click/report/training trajectories.

Scope

• Feature store from events + training_assignments + risk_score_history.
• Train a per-tenant gradient-boosted model predicting click-probability for the next 30 days.
• Calibrated probability surfaced as a percentile.
• Drift monitoring and per-tenant feedback loop.

Acceptance

• Model AUC > 0.75 on hold-out for tenants with >1k events.
• Daily retrain job + per-tenant report.