During design and development the AI designer should apply a relevance criterion to identify the features the network relies on, which are embedded in the trained model and not readable from its structure; the symbolic or relational link from inputs to outputs over a domain shows how much each input drives each output. The result is confirmed either manually, by an expert whose justification can be added to the evaluation report, or automatically, in which case a clearly stated relevance target for the data should underpin the evaluation, together with an explicit method for measuring correspondence between measured and target relevance, a threshold for sufficient correspondence, and the target itself should be provided. The identity and competence of the confirming person or the source of the target can be recorded for traceability.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.