Robustness is the system's ability to keep the level of performance its developers intended under any circumstances, including external or harsh conditions, and it can encompass resilience and reliability; proper operation can bear on the safety of stakeholders. For a machine learning system it shows as generalisation to noisy inputs and the absence of overfitting, achievable for instance by training on large datasets that include noisy samples. Robustness properties show whether performance on atypical or dissimilar inputs is comparable with performance on expected inputs; outputs must stay within a range that is acceptable, consistent or effective, even when they are not ideal. For a regression model it means acceptable response amplitude on any valid input; for a classifier it means no new classification errors appear when inputs move within a range of the typical ones.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.