Loading…
Glossary term
Watching a deployed model's performance and inputs to catch problems.
Once deployed, a model is watched for changes: shifts in input data, drops in accuracy, rising errors, slow responses or unusual outputs. Alerts tell the team when to investigate or retrain.
Where labels arrive late, proxies such as input statistics stand in for accuracy.
Models degrade silently as the world changes, so monitoring is how problems are caught early.