Honest framing. Every image below is a real, biopsy/dermatologist-labelled lesion from the held-out test split the model never trained on. The malignant/benign call, the class probabilities and the attention map are actual output — and the inference runs on Newton's own Rust engine (newton-derm-engine, a hand-written ViT forward on candle), which reproduces the reference network to a max probability difference of 1×10^-6 across all cases. The weights are a pretrained ViT (training Newton's own is the next milestone); the engine computing them is sovereign — no PyTorch at inference. Decision-support, not a diagnosis.