UK MoD JSP 936 Dependable AI in Defence Part 1: Directive (V1.1, 2024) 76-78: Set and demonstrate performance targets for the operating context, reconsider inherited targets, and define responses to out-of-design inputs

What does 76-78 require?

Performance targets for AI components, especially reliability targets, are set out clearly for the operating context and proven to a confidence level matched to the risk of failure. Where AI complements or replaces an existing software, human or combined approach, existing targets are reviewed for continuing acceptability, and where acceptable a Globally At Least Equivalent argument may be used. Appropriate responses to reasonably expected inputs outside the intended design are defined and demonstrated.