A sensitivity criterion expresses the property over a domain and is checked with formal methods. It must at least define the domain on which it was measured and the sensitivity thresholds to be checked, usually as a closed-form variation threshold over a given variation of the inputs. It may be used to compare architectures or trained models provided the networks perform the same task, the same domain is used and the same objective is proven. It is especially suited to interpolation and regression tasks, where it often gives a direct proof against ground truth across the domain.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.