The border is in the image: a clinician-written, training-free rule delineates pigmented skin lesions in a pre-registered blind test of 1,000 dermoscopic images

Background. Automatic delineation of skin lesions is dominated by neural networks. These networks need thousands of expert tracings to train, cannot explain where they place a border, and return a contour even when they fail. Primary care in low-resource settings needs a method that a clinician can state, audit and trust to say "I cannot measure this one". Methods. A nurse stated the rule for the tool in clinical language. It locates the largest region darker than the patient's own skin. It places the border where the colour becomes more similar to that skin than to the lesion core. When its conditions are not met, it refuses to measure and states the cause. The rule was refined on the 100-image validation set of the ISIC 2018 lesion-segmentation challenge (Task 1). The code was then frozen by SHA-256 hash, together with three comparators and the predictions. Finally, it was run once on the 1,000-image challenge test set, which had never been processed by any version of the method. The primary outcome was the median per-image Dice coefficient against the expert mask, with refusals scored as 0. Results. The median Dice was 0.882. The comparators scored 0.695 (Otsu thresholding), 0.661 (centred circle of 30 % area) and 0.598 (centred circle of 50 % diameter). The paired median difference against Otsu was +0.060 (95 % CI +0.047 to +0.077). The tool refused 151 images (15.1 %), each with a written cause; on the 849 accepted images the median Dice was 0.910. The author's pre-registered prediction was 0.88. On the challenge's official metric (thresholded Jaccard index) the tool scored 0.544, below the 0.802 of the best challenge entry. Before the rule was written, the wound-delineation engine it descends from scored 0.36 on the 2,584 training images, losing to the centred circle. Conclusions. A short, clinician-written and training-free rule, frozen before a blind test, delineated the typical pigmented lesion close to expert agreement. It outperformed classical thresholding and declared its failures. It does not match neural networks on the challenge metric, and we report that gap explicitly.