Joint MRI Acquisition Patterns Change Model Separation in Brain-Tumor Segmentation

Brain-tumor MRI sequences arrive in workflow bundles: availability of T1, T1Gd, T2, and FLAIR is jointly structured, whereas standard incomplete-modality tests often hide each sequence independently. We introduce observability conclusion sensitivity (OCS), the change in a fixed model pair's performance margin between two mask distributions, to measure whether the evidence separating methods survives that benchmark choice. Our evaluation freezes a joint prior estimated from 9,842 Centers A/B workflow logs, shares patient and mask draws within comparisons, separates training from evaluation exposure, and instantiates the test with ProtocolProbe, a Swin UNETR probe combining acquisition-state tokens, protocol-conditioned modality masking, and subset consistency. Against mmFormer, ProtocolProbe's margin increases from 2.4 Dice under Rand(0.5) to 4.3 under Protocol, giving OCS = $1.9$ (paired 95% CI [0.9, 3.0], $p=0.001$). After mmFormer receives the same protocol exposure, OCS remains 1.0 Dice (95% CI [0.3, 1.7]); a fixed-backbone ID-only control yields OCS = $0.1$. The response spans 0.7–1.3 Dice across eight separately frozen prior constructions, while external performance margins over protocol-aware mmFormer are 2.4 Dice at both development-isolated Centers C and D. Thus, the full joint mask distribution is a claim-bearing part of an incomplete-modality benchmark: freezing it across methods and reporting its paired margin response reveals comparative evidence that a single i.i.d. dropout score conceals.