Seal vision holdouts across content equivalence

Vision-model holdouts should be sealed across source identity, normalized-raster equality, and perceptual equivalence before inference, because exact byte hashes alone do not detect re-encoded copies.

A reusable selection method is to bind every known training and review registry; reject matching source and normalized hashes, record/capture identities, and close perceptual hashes; require distinct capture groups; choose cases deterministically without model outcomes; and seal the cohort before any provider or model call. If later eligibility depends on an original-image baseline, keep unreadable cases in the fixed selected-source count while excluding them from the smaller comparison denominator without replacement.

This reduces benchmark leakage when images may have been renamed, resized, or recompressed. Its limit is provenance coverage: passing these checks does not prove physical-subject novelty across untracked sources.