Drift monitoring
Watching whether the data, the model's behavior, or the decision environment has moved far enough from the approved baseline to require a review.
Category
What keeps it honest
Related terms
Drift monitoring compares what a system is seeing and doing now against the conditions it was evaluated and approved under. When the gap opens far enough, a change decision is triggered — not by the system deciding for itself, but by a threshold that routes to an owner.
Three things drift separately
Input drift: the population changed. New segments, new channels, new document formats, a seasonal shift that the training window never contained.
Behavior drift: the model's outputs changed relative to the same inputs — usually because something underneath was updated: a base model version, an embedding model, a retrieval index, a prompt.
Environment drift: the decision changed around the model. New policy, new limits, a new product, a regulatory change. The model is doing exactly what it was approved to do, and that is no longer the right thing.
What a threshold breach is
It is the start of an investigation, not a verdict and not an automatic retrain. The breach produces a case: what moved, since when, what downstream decisions it touched, what the data-quality evidence says, and what the options are. A model owner, independent validation where the institution has it, and the governance authority decide from there.
Why it is an operating requirement
A model that is never monitored is not being operated; it is being hoped over. Monitoring is what makes the difference between a system that ages and one that is maintained.
Often confused with
How it differs from similar terms.
| drift monitoring | Compared with | What's different |
|---|---|---|
| Drift monitoring | Evaluation | Evaluation proves a system meets its gates before release. Monitoring watches whether it still does. |
Want these terms worked through on one of your own decisions? Start with a Decision Blueprint