Equal Weight per Instance Does Not Pay Alone: A Blob-Loss Auxiliary on Full-Resolution Cityscapes Buys Thin-Object IoU with Pedestrian Recall, and Only Pays as an Expert of a Mixture-of-Experts

We report a pre-registered, controlled evaluation of the blob loss of Kofler et al. [IPMI 2023] — an auxiliary term that gives every ground-truth instance the same weight regardless of its pixel count — ported to exclusive multi-class semantic segmentation at the native Cityscapes resolution (1024×2048, 19 classes, softmax regime). Arm G (CE + Dice + 0.5·blob) is compared against its exact paired reference B (CE + Dice): same ConvNeXt-V2-Base + UPerNet architecture, same 160-epoch recipe, same augmentation distribution, three shared seeds (42, 123, 456) — only the loss changes. The pre-registered primary endpoint is null: dataset-level official mIoU on the shared 500-image holdout, Δ(G−B) = +0.170 pt, 95 % CI [−0.233 ; +0.551], two-sided paired image-bootstrap p = 0.4032 (B = 10 000 replicates). Equal weight per instance does not improve global mIoU in this regime. What the term actually does is a measured trade-off. It buys thin-object pixel coverage — traffic light +0.98, pole +0.58, bicycle +0.57 IoU vs B (Holm-significant within the 19-class family), truck +3.78 (CI excluding zero), Boundary F1 (3 px) +0.63 — and it pays in pedestrian instance integrity: strict pedestrian recall −4.34, small-instance stratum (T1) −5.65, crowd-group instances −5.74, pedestrian pixel precision −2.70 (all vs paired B, Holm = 0), and a general fragmentation of the masks: connected components rise from 615.1 to 933.9 per image (×1.52), up in 16 of 19 classes against the paired arm B (14 Holm-significant rises; the only material decrease is terrain, −1.8) and in 19 of 19 against the control arm (person masks ×2.2) — a new per-class decomposition that revises the working mechanistic hypothesis of the program (the term does not prune small components; it creates holes). On the program's pre-registered versatility criterion (36 endpoints × 13 arms), G is a dominated specialist: mean percentile 41.9, worst rank 13/13, maximal damage −6.76 pt, 13 significant losses against 4 significant gains. The same verdict holds on a second dataset and a second probability regime: on BRATS 2023 (region-based sigmoid, MedNeXt, nnU-Net recipe, 5-fold CV, n = 1 196), the blob arm alone scored −0.00376 Dice vs baseline, yet as expert 3 of the winning MoE-V3 it contributed +0.00566 Dice, first of 29 arms [DOI 10.5281/zenodo.22903668]. On the Cityscapes program plateau, the expert-mixture arm initialised from these specialists (MoE-V3-CS, master table P3.14) points the same way, as context: ΔmIoU +0.45 pt [+0.11 ; +0.82], p = 0.0066 vs its recipe-paired control, maximal damage −0.53 pt, small-instance recall T1 +0.80, traffic light not degraded — the profile of an arm that puts G's thin-object speciality to use without carrying its damage. The dedicated analysis of that mixture arm (routing diagnostics, multiplicity families) is not the subject of this paper. Contributions. (1) A pre-registered null primary for equal-weight-per-instance auxiliary loss in full-resolution exclusive-softmax segmentation, reported as such. (2) A complete trade-off diagnostic: per-class IoU, seven business-metric families on pedestrian instances, and a new per-class connected-component decomposition (933.9 vs 615.1 components/image) that contradicts the initially hypothesised compaction mechanism — the paper reports what the tables show. (3) An exact port of Kofler's algebra (eq. 1) to the exclusive-softmax regime with precomputed instance packs: +1–2 % epoch time against +870 s/epoch for the naive path, naive↔︎precomputed parity locked by unit tests, and the horizontal-flip alignment pitfall documented. (4) Evidence, on two datasets and two probability regimes, that the term pays as an initialised expert of a mixture while null-or-harmful alone. (5) Public release of code, configs, tables, figures and regeneration scripts.