Parkinson’s disease (PD) is associated with alterations in gait symmetry and plantar loading that can be examined using vertical ground reaction force (VGRF) recordings. This study presents a confidence-stratified, physiology-anchored multiple-instance learning framework with concept-bottleneck-inspired pathways (implementation identifier: DRO-PAS-MIL-CBM; hereafter, PAS-MIL) for retrospective session-level PD-versus-control classification. Each gait session is represented as a bag of temporal windows. Eight predefined bilateral signal descriptors are combined with eight learned latent temporal dimensions, aggregated through attention-based pooling, and processed by concept-guided, prototype, anchor-only, and static-feature expert pathways. The evaluation used five-fold person-grouped cross-validation on 306 sessions from 165 participants in the PhysioNet Gait in Parkinson’s Disease database.Inner person-grouped out-of-fold ExtraTrees probabilities were used to construct the confidence strata and distillation targets. PAS-MIL achieved a pooled session-level area under the receiver operating characteristic curve of 0.771, average precision of 0.890, and a mean fold AUC of $0.826\pm 0.041$. Relevance analysis identified C05 (asymmetry variability) and C08 (bilateral change mismatch) as the highest-weighted predefined physiological anchor descriptors. Protocol-stratified sensitivity analysis showed variation across the three source sub-studies, with AUCs ranging from 0.740 to 0.790. Probability calibration remained suboptimal after temperature scaling (mean per-fold ECE, $0.291\pm 0.042$). The results demonstrate the feasibility of integrating physiology-informed descriptors, temporal representation learning, and session-level aggregation. The study is a retrospective proof of concept and does not establish external robustness or clinical deployment readiness.