arXiv:2610.40710v1 Announce Type: new Abstract: Automated fetal head circumference (HC) measurement reaches near-expert accuracy on the HC18 benchmark, but published models are trained and evaluated on hospital-grade equipment. We quantify what happens when such a model is carried to a low-cost handheld probe. A lightweight segmentation network (CSM, 0.13M parameters) was trained on HC18 across five independent runs and applied without modification to images from a $1,350 handheld probe (SonoQue C3), acquired at the same visit as the reference hospital scan in 18 patients. In-domain accuracy was 2.34 +/- 0.43 mm; on handheld images the same model gave 16.12 mm against the hospital record, a 6.9-fold degradation. Input preprocessing, a choice rarely reported, accounted for about a third of that gap: direct resizing gave 49.45 mm against 16.12 mm for an aspect-preserving pipeline. A leave-one-out data-efficiency curve showed that fine-tuning on two images significantly worsened accuracy, and that from four images upward error fell by 0.75 mm per doubling of the local dataset (p = 0.0045), although no training-set size up to 27 images significantly improved on the unmodified model. Much of the residual gap belongs to the reference rather than the model: scored against a manual tracing of the same image the model was already within 10.48 mm, and its disagreement with sonographer calipers on the same frame was a near-constant offset, negative in 10 of 10 cases. A single calibration constant estimated leave-one-out reduced that disagreement from 8.12 mm to 3.49 mm, reaching the intraobserver reproducibility reported for experienced sonographers, but did not improve agreement with the hospital record. Cross-device evaluations should therefore report the input preprocessing used and characterise the precision of their reference standard.