What counts as finding a cryptic pocket?

A cryptic pocket is a druggable site absent from deposited structures, open only when the protein transiently samples a higher-energy conformation. Many methods now predict such sites; far fewer address how to confirm one, and the field does not agree on what counts as finding one computationally. This paper asks what it takes to call a cryptic pocket found. (1) Most cryptic pockets cannot be tested at all. The validation experiment the field runs, thiol labeling of a buried cysteine the pocket exposes, can reach only 13.0% of CryptoBench's 2,267 annotated cryptic pockets. (2) That makes finding pockets a screening problem, we make it cheap. We pretrain a 13.2 M-parameter encoder, using no cryptic labels, to reproduce frozen ESM-2 650 M from local structure. An uncertainty head on that representation then ranks residues as cryptic zero-shot at AUC 0.771. A supervised detector on the same representation, with no PLM, matches CryptoBench's own 3 B-parameter baseline (pLM-NN) at 153$\times$ fewer inference parameters, and shortlists testable pockets at 39.1%, cross-validated over held-out proteins, against a 0.39% chance rate, roughly double what P2Rank reaches in the same pipeline. On CryptoBench it more than doubles the AUPRC of PocketMiner, 0.406 against 0.19. (3) But scoring a cryptic pocket is not a settled convention. The field's centroid-distance criterion is applied anywhere from 4-12 Å, without a chance rate, meaning a score cannot be separated from the criterion's own permissiveness. Moreover the two models just compared were trained on different definitions of a cryptic pocket, apo-holo ligand binding against MD opening.