Validation & Uncertainty Quantification

Validation activities should be used to demonstrate that the model is fit for its intended Context of Use. Credibility evidence should be generated proportionate to risk, addressing assumptions, uncertainties, and limitations and should be used to determine whether the model is fit to meet the Question of Interest within the defined Context of Use.

Within this level, the model training should be described, involving the estimation or optimisation of model parameters using the defined datasets. For data-driven components, this includes training, calibration, and tuning; for mechanistic models, parameter estimation and calibration; and for hybrid models, coordinated training of integrated elements. Training should be performed under controlled conditions, with clear separation of training and evaluation data where applicable, and with attention to avoiding overfitting and unintended bias.

Furthermore, the model evaluation process should be described, where the trained model should perform as required for its Context of Use. This includes evaluation of predictive performance, robustness, uncertainty, and limitations using appropriate quantitative and qualitative metrics. Results should be interpreted against predefined acceptance criteria derived from the risk assessment, supporting an evidence-based judgement of model credibility. Outputs from this stage should inform subsequent credibility review, regulatory justification, and the decision on model fitness for purpose.