Drift Detection

Drift occurs when a model’s output behavior changes from the baseline established when the model was registered and approved. Drift can happen due to model updates, prompt changes, context shifts, or data distribution changes.

Setting a baseline

When you register an AI system in GOVERN, you can establish a behavioral baseline by running a representative set of inferences through GOVERN. GOVERN records the distribution of assessment scores, scorer values, and output characteristics as the baseline.

How drift is detected

GOVERN continuously compares incoming assessment distributions against the stored baseline. When the distribution diverges beyond your configured drift threshold, GOVERN raises a drift detection alert.

Drift signals GOVERN monitors

• Assessment score distribution shift
• Violation rate change
• Scorer value distribution change (per scorer)
• Output length distribution shift
• Response latency change (proxy for model substitution)

Drift thresholds

Configure drift sensitivity in Policies → [Your Policy] → Drift Settings. A lower threshold means more sensitive detection (more alerts, potentially more false positives). A higher threshold means only significant behavioral shifts trigger alerts.

Responding to drift

Drift alerts link to a drift detail view showing the before/after distribution comparison. If drift is confirmed, open a remediation to investigate the root cause: model update, prompt change, or data shift.