Abstract
<jats:p>This article proposes persistent homology as a model-agnostic audit lens for vision–language medical systems whose fluent reports can mask structural inconsistencies. An end-to-end pipeline computes cubical persistence on reference images and model-mediated outputs, derives Betti-curve signatures on a shared filtration grid, and measures drift via bottleneck and Wasserstein distances in H0 and H1. Conventional baselines (reconstruction, segmentation, and semantic agreement) establish nominal adequacy, while topology reveals silent failures such as spurious holes, merged components, or fragmented regions. Topological signals are aggregated into a diagnostic index for ranking cases, setting alarms, and tracking stability. Outputs include figures, tables, and evaluation scripts.</jats:p>