New Framework for Counterfactual Image Auditing and Feature Bounds

The framework uses common witness grades and witness nerves to identify failures where local image editing constraints fail to satisfy global latent explanations. It provides sharp partial-identification bounds for image features and uses Helly-type arguments to generate incompatibility certificates for quasiconvex losses.