The scenario
A lending platform must log loan applications, model predictions, and actual outcomes to evaluate bias across protected attributes over time.
What is in this drawing
Read the decisions behind it.
01
Separate ApplicantAttribute join table to track multiple protected attributes per applicant
02
Decouple ModelVersion and MonitoringRun so each deployed model version can be audited independently
03
Store BiasMetric thresholds and pass flags alongside Alert records for reproducible detection
Use when building a monitoring data lake for ML models that need bias auditing, drift tracking, or fairness reporting.
More like this