Leaf identity and visual group splitting in potato disease classification

Image similarity does not, by itself, establish repeated specimen identity or duplicate leakage. We reevaluate potato leaf classification using 240 completed runs spanning 4 architectures, 2 datasets, 2 holdout strategies, and 15 seeds. The laboratory dataset contains 2,152 PlantVillage images mapped to 538 published leaf identities; the field dataset contains 1,296 images from a single source, classified as Healthy, Fungi, or Phytophthora. The laboratory grouped holdout preserved leaf identities, whereas the field grouped holdout preserved DINOv2 visual clusters. Subset sizes and class allocations were constrained. Accuracy, balanced accuracy, macro F1, per-class metrics, and relative split gaps were calculated from saved predictions. Against published laboratory identities, DINOv2 clustering at a cosine-distance threshold of 0.10 yielded a pair precision of 3.89%, pair recall of 49.91%, and identity purity of 38.34%. The mean proportion of laboratory test image whose identities were also present in the training set was 97.13% under random holdout and zero under identity-grouped holdout. Random-minus-grouped balanced accuracy gaps range from 0.27 to 4.03% points in the laboratory dataset and from 5.31 to 9.54% points in the field. Seventeen of the 24 comparisons met the specified false discovery rate criterion, including all 12 field comparisons. No exploratory ViT vs.-mean-CNN interaction contrast survives its separate six-test adjustment. The results establish sensitivity to the evaluated partition protocols, not a causal decomposition of duplicate effects or a general difference between CNN and transformers. Validation of near-duplicate identity in the field dataset remains unavailable, limiting claims about leakage.