Skip to main content

Model Promotion Policy

Define objective promotion gates before publishing model aliases.

Required Inputs​

  • Latest training report (.md/.html)
  • Holdout and walk-forward metrics
  • Split metadata and dataset provenance

Suggested Gates​

  • Holdout AUC above baseline threshold.
  • Walk-forward mean quality above floor.
  • Acceptable variance across folds.
  • Stable threshold policy and non-trivial coverage.

Promotion Steps​

  1. Validate metrics against policy.
  2. Snapshot current production aliases.
  3. Publish new aliases (<Strategy>.joblib or ensemble aliases).
  4. Enable monitoring for post-promotion drift.

Rollback Rules​

  • Roll back if critical quality or runtime health alerts fire.
  • Keep previous alias set immutable for quick restore.