Evidence-Governed Federated Learning

Federated Learning (FL) often links statistical volume to aggregation influence. However, evidential support from learned contribution does not have to be determined by the client's data amount. This work introduces Evidence-Governed Federated Learning (EGFL), which includes a six-dimensional validation-evidence profile and client-derived trust. It avoids including sample size as an input before statistical volume and trust are combined to determine the influence of aggregation. It follows 295 prespecified tasks across two heterogeneous multi-hospital ecosystems, independently generated recovery experiments, matched reliability failures, five federated methods, and an aggregation rule ablation over five random seeds. EGFL is not a predictive optimisation substitute, as it never outperforms the comparators in absolute predictive performance. EGFL's degradation was consistently greater at the tested high-volume client than at the small-volume client under the same configured combined reliability corruption. Compared to FedAvg, FedProx, and FedBN, the framework better attenuates the excessively high-volume failure burden across five of the six metric levels, with uncertainty confined to F1. SCAFFOLD provides the best performance as a counterexample to universal attenuation. Aggregation ablations reveal weak performance due to low trust relative to volume-containing rules in the tested setting. These findings position the EGFL framework as a mechanism that keeps evidence-derived trust and statistical volume distinct yet jointly determines client influence, thereby allowing for the reduction of influence due to reliability deterioration otherwise conferred by data volume without eroding statistical contribution.