Lacuna discovers cryptic binding pockets in proteins by generating a conformational ensemble, detecting pockets per conformer, clustering across the ensemble, and ranking the resulting sites with a learned model. On CryptoBench's designated test fold (n=178) it recovers 66.3% of known cryptic sites in its top 5 with the protein-language-model ranker, under a size-robust Jaccard criterion, against 63.5% for P2Rank, 61.8% for IF-SitePred and 43.8% for fpocket. The difference from P2Rank is not statistically separable. Runs on CPU with an Anisotropic Network Model backend.