For a classifier, robustness depends on how well the model separates classes. When designing one, a sensitivity criterion is used to find which classes are more or less separated: the sensitivity analysis must work on domains constructed around individual test data points, and the spread of attribute values should be widened gradually to see which classes start to overlap, overlap meaning that the output for one class begins to exceed that of another, stopping when every class overlaps every other. The order in which overlaps appear and the domain size at which they begin drive action on the training data or the architecture to improve separation.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.