Data for machine learning is shown to bear on the behaviour intended within the ODD, complete across the input space with real-world distributions, balanced from the model's perspective (for example over-representing rare classes), and accurate as a reflection of the real application including sensor characteristics and labelling. Separate training, test and verification data are produced from context-relevant sources and kept independent; subsets split from one dataset are verified as sharing its characteristics; and where possible verification data is collected separately and spans the ODD including edge cases inside, on and beyond its boundaries.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.