Duplicate leakage in a public conjunctival pallor benchmark, and an honest baseline for image-based anaemia screening

CP-AnemiC is a public dataset of 710 conjunctiva photographs of Ghanaian children aged 6 to 59 months, each paired with a laboratory haemoglobin value, and it is now used as a benchmark for image-based anaemia screening. We show that only 383 of the 710 records are distinct photographs. The duplication takes three forms, and each one is invisible to the check that catches the previous one: 212 byte-identical copies; 19 re-encoded copies whose bytes differ; and 96 copies shifted or re-cropped by a few pixels, with different hand-drawn masks, which neither file hashing nor pixel comparison finds. The copies are not harmless. At the final grain, 116 groups carry conflicting haemoglobin labels for the same photograph, and 125 groups are attributed to more than one hospital, so grouping folds by collection site, the usual safeguard, does not contain them. A random split of the raw records inflates AUROC by +0.176 (95% CI +0.120 to +0.232), and an analysis that hashes files and groups folds by site still leaves roughly two-fifths of that inflation in place. With all three forms removed and folds grouped by site, twelve interpretable colour features reach AUROC 0.706 (95% CI 0.653 to 0.757) for WHO anaemia, +0.182 over demographics alone. Bland-Altman limits of agreement span 7.8 g/dL, so the model is a triage screen and not a haemoglobin meter. We release the shift-invariant detector and the full analysis, and recommend that future work on this benchmark state which forms of duplication it tested for.