Precision Pays: A Fixed-Weight Boundary Loss Anchors the Precision End of a Recall–Precision Axis and Wins as the Consensus Veto under the Official BraTS-2023 Metrics

Preprint (Paper 2 of the BRATS auxiliary-loss program). Does adding the Kervadec boundary loss (fixed weight lambda = 0.05, acting directly on the logits, no auxiliary head) to a strong MedNeXt/nnU-Net glioma backbone improve adult glioma segmentation? Evaluated under the complete official BraTS-2023 metric set over a 5-fold held-out cross-validation on the 1196-case study set — the unified BraTS-2023 GLI download holds 1251 cases, and the 55 that never entered the nnU-Net dataset are listed one by one (trigger, git date of the triggering entry, and a fresh re-measurement of their file integrity against 55 included controls: 0 unreadable file, 0 NaN/Inf, 0 out-of-set label, 0 divergent affine) in analysis/DATASET_EXCLUSIONS.md. Against the baseline the boundary loss is not significant on any pre-specified primary or region-averaged official metric (lesion-wise Dice delta = +0.000, Holm p = 0.34; lesion-wise HD95 delta = -0.58 mm, p = 0.31), a null confirmed by a 3-seed multiseed check (paired t-test p = 0.42). Its only robust distinctive property is being the precision-oriented mirror of a signed-distance (SDT) auxiliary head — significantly more specific, fewer spurious tumor-core/enhancing lesions, lower recall — which makes it an effective connected-component consensus veto: used to veto the SDT model it removes about half of that model's false-positive lesions at negligible recall cost (lesion-wise HD95 -9.95 mm, p = 2.3e-20 over the 5-fold CV), and used as the primary model with a baseline veto it reproduces the companion paper's consensus gain (lesion-wise Dice WT +3.89 pp, p = 0.005, 3 seeds x fold 0). Reported in full as a transparent negative result, with a weight-selection robustness check ruling out mistuning. The archive bundles the manuscript (EN + FR, with both rendered PDFs), the reproducibility code, the exact result JSONs behind every reported number (including the patient-level export that carries the Figure 3-5 counts and the training-log export that carries the multiseed validation Dice), the tables, figures, the measured dataset-provenance artefacts and the verified reference list. All materials are released together under CC-BY-4.0 as a single publication record.