robustness-eval

A small, referenceable corruption-robustness benchmark for image classifiers (ImageNet-C / CIFAR-10-C style). Train clean, grade under corruption × severity, report the robustness gap. Agent-drivable --json CLI.