Where it shines

Guaranteed recall for screening & retrieval

When the deliverable is a shortlist, coverage is recall, and conformal prediction certifies it.

Drug-candidate triage, document retrieval, fraud review: the product is a shortlist, and the thing that must not happen is dropping the real hit. That is a recall guarantee, and it is coverage in disguise. Take a calibration set of genuine hits, look at the scores your model gives them, and set the keep-threshold \(t\) at the conformal lower quantile of those scores. Keep everything scoring \(\ge t\). By exchangeability a new genuine hit survives the cut at least \(1-\alpha\) of the time, a distribution-free recall guarantee that does not care how the scores are distributed. Drag model separability: the guarantee never moves, but a sharper model lets you hit the same recall with a far shorter list.

The needles (relevant items) and the haystack (irrelevant), by model score. The conformal threshold is placed so a new needle clears it with probability \(1-\alpha\); everything to its right is the returned shortlist. Recall is guaranteed; how much haystack comes along, the precision and list length, is what the model controls.