The institution schedules tests of data quality and functionality covering accuracy, robustness, stability and where necessary bias, with experts from the area of application supplying the questions and expectations and performance indicators defined in advance to judge whether the application meets its goals; it defines thresholds or other validation methods (sampling, backtesting, predefined test cases, benchmarking) for regular checks of output correctness and quality, monitors changes in input data to detect and treat data drift so models stay applicable, analyses the cases where users ignored or changed the output as a signal of weakness, and decides in advance how exceptions are recognised and handled, with fallback mechanisms should the AI drift from its purpose.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.