Rules + ML Fraud Decision Path

Hard rules, model scoring, policy bands, outcome log, and champion/challenger retraining

Rules + ML Fraud Decision Path Hard rules, model scoring, policy bands, outcome log, and champion/challenger retraining 01 / Input 02 / Screen 03 / Score 04 / Decide 05 / Learn Transaction · auth event · 01 / Input · inbound Transaction auth event inbound Hard Rules · block only · 02 / Screen · tier 1 Hard Rules block only tier 1 Features · soft signals · 03 / Score · assembled Features soft signals assembled ML Model · risk score · 03 / Score · tier 2 ML Model risk score tier 2 Policy · band to action · 04 / Decide · approve / step-up / decline Policy band to action approve / step-up / decline Outcome Log · features + decision · 05 / Learn · immutable Outcome Log features + decision immutable Trainer · champion / challenger · 05 / Learn · retrain Trainer champion / challenger retrain raw event passed soft signals score + band decision + reason labeled outcomes promoted model Legend primary data policy / PII async batch data store

One Decision Path

  • • Hard rules run first and can only block, never approve
  • • The model owns the fuzzy middle via score bands
  • • Every decision has one unambiguous source

Bands, Not a Threshold

  • • Low risk approves, high risk declines with reason codes
  • • The review band routes to a recoverable step-up
  • • Soft signals pick which step-up challenge to issue

Closed Feedback Loop

  • • Outcome log snapshots features as scored, avoiding skew
  • • Chargebacks and confirmations arrive as delayed labels
  • • Challengers shadow live traffic before promotion

Data-flow diagram • Built with Archify • Create yours ↗ • Hover to trace • R route • Click to focus • +/− zoom • M radar • T theme • E export