SSAT: A reproducible toolkit for task-conditioned spatial sensitivity auditing of vision models

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

SSAT (Spatial Sensitivity Audit Toolkit) is an open-source Python package for reproducible, dataset-scale auditing of how image and video classification models respond to interventions on user-defined spatial or semantic regions. It reports effect sizes alongside evidence grades from matched controls and cross-operator checks, records preprocessing-aware geometry that exposes area confounds, and keeps full model outputs with provenance for inference-free reanalysis of stored interventions. Synthetic, ImageNet-1k, and NTU RGB+D studies demonstrate shortcut recovery, geometry diagnostics, and protocol reuse across CNN and transformer classifiers and modalities; further analyses quantify grade sensitivity to thresholds and control counts, and audit cost scaling.