Glossary
Distribution shift
When the data a model sees in use differs from the data it was trained or calibrated on, so its accuracy and calibration may change.
Checks run on our GPUs in Switzerland. Request payloads are not stored.Security
What Distribution shift means
Distribution shift means the inputs in production do not look like the inputs a model was trained or tuned on. For a guard, that could be new tools, a different agent framework, another language, longer contexts or a different mix of safe and unsafe actions.
Under shift, two things can degrade. Ranking quality (AUROC) may drop, and calibration may drift, so that a threshold which gave a certain precision on the original data gives a different one on yours. Studies such as Ovadia et al. (2019) found that predictive uncertainty generally degrades under dataset shift.
The practical consequence: do not assume published benchmark numbers or default thresholds carry over unchanged. Evaluate the guard on a sample of your own labelled actions, fit thresholds per deployment, and re-check when your tools or agents change.
Related terms
How Distribution shift connects to the rest of the vocabulary.