This is a marimo notebook walking through Salvatore, Fort & Ganguli's recent claim that adversarial examples are geometrically inevitable: every concept a network learns lives on a "perceptual manifold" that fills almost all of pixel space, so any image is one short pixel-walk from being misclassified. Below, I reproduce the Appendix E geometry, sample CLIP's perceptual manifold for "a photo of a cat," run a real ε-bounded adversarial flip, and then test (and fail to confirm) one alternative explanation that traces the effect to text-side bias. The notebook is pretty heavy, so it might timeout on Marimo's compute. If so, downloading and running it locally should work!