CLIP style-classification accuracy drops from 0.87 to 0.77 once the model cannot recognize the artist.

Standard evaluation of frozen-embedding style classification uses random splits where works by the same artist appear on both sides (arXiv 2608.14435). Under an artist-disjoint protocol on 320 paintings across four twentieth-century movements, 5-NN accuracy falls ten points, and unevenly: Impressionism and Cubism barely move, Surrealism drops twenty. The pattern holds across four encoders including a vision-only self-supervised model, placing the effect in visual structure rather than language. A clean template for leakage-by-entity in any embedding benchmark.