Prenatal ultrasound biometry must return a head circumference (HC) when the image supports it and refuse otherwise. Selective prediction trades off coverage against error; its finite‑sample guarantees only bound the rate of gross errors, not their magnitude. We audited a tool‑grounded biometry pipeline on two pre‑registered four‑way splits of the 999 labelled HC18 images, retraining the segmenter three times on each split. Calibrating the contour‑completeness threshold with exact Clopper‑Pearson risk control produced a certificate indistinguishable from a hand‑set floor, yet the pipeline still emitted a 271.3 mm error: a small geometrically complete ring fitted to an image whose true HC is 324.6 mm. The candidate scored 0.94 completeness, so the score missed the failure; one outlier among roughly 150 answers kept the certified risk bound (α = 0.10) intact. A second untouched split reproduced the mechanism on a different image (233.3 mm). Constraining the admissibility region—the anatomical range where the measurand is defined—removes the whole error family: raising the lower bound from 50 mm to 55 mm costs only 0.4 percentage points of coverage but cuts the worst‑case error to 11.4 mm and 11.1 mm, whereas suppressing it via the completeness threshold costs 19.0 and 12.1 points respectively. At an anatomically defined 70 mm lower bound, frozen tests yield MAE of 1.91 mm and 1.81 mm.
Review: The risk‑control framework guarantees only the error rate, leaving room for isolated large mistakes. Incorporating anatomical priors into the admissible range effectively reduces the maximum error but reduces coverage. This audit highlights the limitation of selective prediction in medical imaging and the need for domain constraints to achieve reliable automated measurements.