Tracing high transductive cohort AUC to same-site supervision in a site-aware population GNN for multisite fMRI

Population-graph models can exploit cohort-level context, which complicates the interpretation of high multisite neuroimaging performance. We asked which information pathways account for a previously reported high transductive cohort AUC in multisite autism fMRI. Under a frozen cohort (canonical ABIDE-I, 871 subjects, 20 sites) and a fixed evaluation protocol, we applied controlled graph, feature, supervision, and architecture interventions to a clean-room re-implementation, replicated across two preprocessing pipelines (C-PAC and NIAK). Cohort out-of-fold AUC was approximately 0.94 and depended on site-linked edges: a site-only graph matched the full model (0.948 vs. 0.941). The gain required same-site supervision (masking it reduced AUC to 0.480), was not explained by the supervision budget, and, among architecture-matched heads, was observed only with the evaluated sex-heterogeneous dual-channel head, whereas a canonical topology-only Parisot-GCN did not show the same pattern. Under leave-one-site-out evaluation, performance fell to near-chance AUC (0.522 for C-PAC and 0.532 for NIAK), whereas an imaging-only reference remained at 0.653 and 0.588, respectively. The high cohort AUC therefore depends on same-site supervision through the evaluated transductive graph under this protocol, whereas it does not translate to unseen-site discrimination.