Where high-level requirements implemented through AI decompose directly into low-level requirements they are documented and traced as for traditional software; where behaviours are learned and not directly decomposable, the combination of training algorithm and data requirements is demonstrated to meet the intent of the high-level requirement. High-level requirements clearly articulate the operating context including training and test data distributions; where data may be sparse or missing, requirements and guidance for mitigation across the lifecycle are provided; data requirements for machine learning state how set-aside data is curated for independent verification; and every requirement says how verification and validation will show it has been met with meaningful measures of acceptable performance in test and operational environments.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.