When documenting the models used, the organization should consider: the data (training, validation, test) and algorithms used to build them; the training methods; confirmation that no data is repeated across the training, test and validation sets; the steps for training, testing and validating several models; the criteria and steps for feature selection; the model selection criteria; the metrics used to evaluate model performance (such as accuracy, the receiver operating characteristic, the confusion matrix, mean squared error or mean distance to cluster centre); how well the chosen models generalize to production data; whether they preserve or create unwanted bias or other harm; whether their outputs contain or create personal data; robustness; detecting and correcting data drift (for example retraining when production data diverges from the original training data); retraining criteria and process; the effects of continuous learning where used; and computational and environmental considerations. Models the organization fine-tunes are reviewed against this list for the new data and scenarios, with the provider's information consulted first; for imported pre-trained models the selection, evaluation, bias and personal data items and the testing parts of the data, separation and multi-model items are reconsidered for the new data and scenarios.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.