Conformal monitoring is the practice of turning a model's outputs into calibrated evidence of change without assuming how the data is distributed.
Also known as: conformal prediction for monitoring
Conformal monitoring is the use of conformal methods to turn raw model outputs into well-calibrated signals you can feed to a change detector, even when the underlying data is messy or non-normal. Conformal methods let you attach honest uncertainty to a prediction without assuming the data follows any particular distribution — you only need past examples to compare against.
For a practitioner, the appeal is robustness: you do not have to hand-fit a distribution to each metric before you can trust its alarm. The calibration comes from your own history.
Paired with e-processes, conformal scores become the input to an anytime-valid detector — distribution-free calibration on the front end, a valid-under-peeking guarantee on the back end.
ValidAnytime turns these ideas into a live alarm you can trust — valid no matter how often you look. Prove it on your own data, free.