Model Cards

A game-theoretic approach to explain the output of any machine learning model. SHAP values quantify the contribution of each feature to a specific prediction, providing the granular, per-decision justification often summarized in a model card's evaluation section.

• Connects feature importance to Shapley values from cooperative game theory
• Unifies several explanation methods, including LIME and DeepLIFT
• Essential for generating the adverse action reason codes required for regulatory compliance

A comprehensive, chronological record of the data, model parameters, decisions, and logic used for a specific transaction. While a model card provides a static, high-level summary of a model's intended use and limitations, an audit trail provides the dynamic, per-transaction evidence required for full traceability.

• Designed for regulatory review and internal governance
• Complements model cards by providing the runtime execution log
• Critical for demonstrating adherence to the stated limitations in a model card

The institutional framework of policies, procedures, and controls that manage the lifecycle of AI models. Model cards serve as a key artifact within this framework, acting as a standardized disclosure document that supports model inventory, risk tiering, and regulatory compliance automation.

• Addresses model risk management (MRM) requirements for audited environments
• Ensures alignment with regulations like the EU AI Act
• Transforms model cards from a documentation exercise into a governance control

A property of models that are inherently simple and understandable due to their structure, such as linear regression, a single decision tree, or an Explainable Boosting Machine (EBM). A model card for an intrinsically interpretable model can make stronger claims about auditability and safety, as the decision logic is directly inspectable without post-hoc approximations.

• Contrasts with post-hoc explainability required for black-box models
• Enables modular, human-readable explanations of predictions
• Simplifies the 'Ethical Considerations' section of a model card

A method that explains a model's decision by identifying the minimal changes to an input instance that would alter the prediction to a desired outcome. In the context of a model card, counterfactuals can be used to illustrate the model's decision boundary and demonstrate fairness by showing what a denied applicant would need to change to be approved.

• Answers the question: 'What would need to be different?'
• A powerful tool for algorithmic recourse and bias detection
• Provides intuitive, human-actionable insights for the 'Limitations' section

A class of interpretable neural networks that first predict a set of human-defined concepts from the input and then use only those concept scores to make the final prediction. A model card for such a model can explicitly list the intermediate concepts used for reasoning, providing a verifiable chain of logic that aligns with human understanding.

• Forces the model to learn a reasoning process aligned with human concepts
• Enables intervention by allowing a human to correct a mispredicted concept
• Makes the model's internal logic directly auditable and documentable