Definition
Plain language
A way to attach a reliable error rate to a system's decisions without assuming anything about the data.
As stated in the literature
A distribution-free framework that calibrates a decision threshold from observed outcomes to guarantee a target error rate; the basis for Conformal Decision Theory's adversarial long-run violation-rate guarantee.
Why it matters: It puts a trustworthy error rate on a system's choices, which is essential when wrong decisions carry real consequences.
For example, a system can be tuned so that no more than five percent of its confident decisions turn out wrong, without assuming the data follows any neat pattern.