Rank Radii Transfer as Quantiles

A cluster is not a determination. Generalized category dis- covery will sort unfamiliar images into clean groups, and the groups are often correct, but a taxonomist reading them still cannot say whether the third group is a new genus of a known family or the first specimen of something nobody has described. We recast novel species discovery as taxonomic placement: name the deepest rank at which a new species can defensibly be attached to the known taxonomy, or refuse. Refusal is not an edge case. Across a 550-species bench- mark built from iNaturalist-2021, only 4 of 45 novel species have a congeneric relative in the reference set. Eighteen attach at family, nineteen at order. A further 100 species come from two clades withheld in full, and those have no ancestor at any rank, so the only correct output is that no attachment exists. The method follows from one property of the embedding. Similarity between species falls monotonically with taxonomic distance, so a rank behaves like a radius, and a dendrogram cut at that radius recovers that rank. The radius does not survive the move to a new clade. Densities differ, and a threshold calibrated on known species merges 88 true genera down to 48. Its quantile does survive. Read as a position in the local similarity distribution rather than as a similarity, the same construction recovers all 88, with no taxon count supplied at any rank, and 86 to 89 genera at a mean genus NMI of 0.973 ±0.004 across five independent clade hold-outs. Placement into the existing taxonomy reaches hierarchical F1 of 0.715 against 0.534 for nearest-neighbour attachment, and cuts mean rank error from 1.511 to 0.467. Then it stops. Six interventions failed to move parent accuracy, because when a species’ own family is absent from the reference set, a sister family in the same order is frequently the closer match. That is a property of what images encode, not of the classifier.