Visually similar prints expose a benchmark validity gap in computational art history

Digitised cultural collections increasingly use visual retrieval to connect related images, yet visual resemblance does not by itself establish a historical relation between physical objects. This distinction is acute for prints: two sheets may preserve the same composition while deriving from different blocks, states, workshops or copying events. The study tests whether a retrieval benchmark can appear successful because it omits such historically consequential alternatives. A diagnostic repeated-design task is compared with a controlled candidate-pool intervention built from 39 Metropolitan Museum of Art woodcut records. Across 240 transformed queries, Pool A contains an intended same-design target and other-design negatives; Pool C changes nothing except adding one catalogue record explicitly marked as a same-design copy. Raw pixels, histograms of oriented gradients and ORB rank the intended target first for all Pool A queries, but only 187, 177 and 237 times in Pool C. The added copy becomes the unique top result 53, 63 and 3 times. An abstaining ORB/HOG gate also accepts a copy for two queries. These are catalogue-class ranking errors, not findings about matrix identity or authenticity. The results show that candidate composition is part of construct validity: a benchmark of design recurrence cannot support a claim about historical sameness unless relation-confusable candidates are represented. A relation-aware evidence architecture is proposed that separates visual discovery, confusability testing and source-bearing historical judgement. The approach gives digital humanities and cultural-heritage projects a reproducible way to test when useful similarity becomes an invalid historical inference.