“NVIDIA's autonomous driving software architecture runs two stacks in parallel: an end-to-end model and a "classical stack" that acts as a safety guardrail to verify the model's outputs.”