Overview
Most teams treat model drift, governance, and infrastructure efficiency as separate problems. In practice, they are the same operational problem expressed through different metrics.
This research proposes a unified view: drift signals should trigger decisions that are simultaneously performance-aware, compliance-aware, and carbon-aware.
Research objective
Build a decision protocol for production ML systems that answers:
- When should we retrain?
- When is rollback safer than retrain?
- Which action minimizes regulatory risk and operational waste?
Proposed operating model
1) Drift detection layer
Use data drift and concept drift indicators as the first trigger:
- population stability index,
- prediction confidence distribution shift,
- segment-level error deltas.
2) Governance layer
Map triggered events to governance policies:
- risk tiering aligned with EU AI Act logic,
- mandatory audit trail for high-impact model changes,
- model card delta log for each intervention.
3) Time-Money-Carbon layer
Before acting, estimate each candidate response:
- Time: expected recovery time to acceptable model quality,
- Money: compute + engineering effort cost,
- Carbon: estimated energy and emissions from retraining or rollback alternatives.
Hypothesis
For many incidents, selective rollback and targeted recalibration can outperform full retraining when evaluated across all three dimensions.
Planned outputs
- A practical triage matrix for drift incidents.
- A lightweight scoring template for intervention decisions.
- Case studies showing where compliance and efficiency objectives conflict, and how to resolve them.