Model Governance & Spec• LightGBM Risk Engine

Model Architecture & Governance

PRODUCTION ACTIVE
Production Model Lineage Pipeline
01 / DATASETIEEE-CIS Fraud590,540 rows
02 / FEATURES492 DimensionsSingle-pass vector
03 / BOOSTERPISTA LightGBMActive Model
04 / CALIBRATIONIsotonic MappingBrier Score: 0.0234
05 / DECISIONINGTri-State RoutingFrictionless Triage

Interactive Policy Threshold Studio

Simulate real-time operational trade-offs between automated pass-through and manual triage load.

Review Threshold (τ_review)25%
5% (Aggressive Review)50% (High Pass-Through)
Block Threshold (τ_block)75%
50% (Strict Hard Rejection)95% (Lenient Block)
Frictionless Approval

84.0%

Zero merchant friction
Manual Triage Load

11.0%

Routed to investigators
Hard Automated Block

5.0%

Instant rejection
Expected Fraud Catch

93.5%

Chargeback suppression

PISTA LightGBM Risk Engine

Calibrated LightGBM GBDT pipeline (v1.0.0)
PRODUCTION ACTIVE

Optimized gradient booster with leaf-wise tree growth, balanced logloss, and integrated Isotonic probability calibrator.

PR-AUC0.5450
ROC-AUC0.9130
Inference< 1.0 ms

Heterogeneous Tri-Blend

Challenger: LightGBM + XGBoost + CatBoost
BENCHMARKED

Three-way stacked blend. Retained for governance validation but bypassed for live transactions due to higher latency overhead.

PR-AUC0.5482
ROC-AUC0.9142
Inference4.20 ms