Model Card

The foundational metadata section identifying the model's version, type (e.g., gradient boosted tree, transformer), development date, and responsible team. This section establishes a unique identifier for lineage tracking and audit trail purposes, ensuring every model instance is unambiguously referenced in governance documentation. It typically includes the model owner, point of contact, and the specific business unit accountable for its performance under SR 11-7 guidance.

A precise, bounded statement defining the specific use case and population for which the model was designed and validated. This section explicitly delineates the operational context—such as 'detecting point-of-sale credit card fraud for U.S. domestic transactions under $10,000'—and equally importantly, declares out-of-scope uses that are unsupported and potentially harmful. This boundary-setting is critical for model risk management and prevents unsafe extrapolation.

A transparent disclosure of the model's performance using standardized quantitative benchmarks, segmented by relevant cohorts. This section reports metrics such as AUROC, precision-recall at fixed thresholds, and false positive rates, often broken down across demographic or transactional segments to surface performance disparities. For fraud models, it includes the Population Stability Index (PSI) to monitor data drift and backtesting results against historical outcomes.

A summary of the data sources, time periods, and sampling strategies used to construct the training and validation datasets. This section discloses the prevalence of the target class (e.g., fraud rate of 0.17%), any synthetic data generation techniques applied to address class imbalance, and the feature engineering approach. It provides essential context for validators assessing conceptual soundness and for identifying potential selection biases embedded in the training corpus.

A candid assessment of the model's potential for disparate impact and the results of fairness evaluations. This section presents the outcomes of disparate impact testing across protected classes, reports adverse impact ratios, and documents any mitigation strategies applied, such as threshold adjustments or feature exclusion. It directly supports compliance with fair lending analysis requirements and the EU AI Act's mandate for fundamental rights impact assessments.

An explicit enumeration of the model's known weaknesses, edge cases, and conditions under which performance is expected to degrade. This includes sensitivity to concept drift (e.g., new fraud typologies), reliance on specific data quality dimensions, and performance degradation on tail-end segments. By pre-declaring limitations, the card enables informed human-in-the-loop override decisions and sets expectations for continuous model evaluation cadences.

A game-theoretic attribution method that decomposes an individual prediction into the additive contribution of each input feature. Model cards frequently include global SHAP summary plots to disclose which features drive decisions across protected demographic cohorts. This enables fair lending analysts to verify that no prohibited basis variables are influencing outcomes, directly supporting disparate impact testing requirements.

A symmetric metric quantifying distributional shift between development and production data. A model card documents the PSI thresholds that trigger review, establishing the quantitative boundary for acceptable drift. When PSI exceeds these thresholds, the card's documented limitations section guides investigators toward the specific features requiring root cause analysis and potential retraining.

A silent degradation mechanism where input feature distributions change in production without an immediate shift in the target outcome. The model card's evaluation benchmarks section establishes the baseline distributions against which drift is measured. Continuous monitoring dashboards reference the card's documented acceptable ranges, automatically flagging when feature statistics deviate beyond the attested boundaries.

A quantitative methodology that identifies facially neutral decision rules which disproportionately affect protected groups. Model cards explicitly disclose adverse impact ratios across segments—such as race, gender, and age—calculated during pre-deployment evaluation. This transparency transforms the card into a regulatory artifact demonstrating proactive compliance with fair lending laws and the EU AI Act's fundamental rights requirements.

A human-interpretable statement identifying the minimal feature changes required to flip an adverse model decision. Model cards often include canonical counterfactual examples for denied transactions, illustrating the decision boundary in plain language. This bridges the gap between opaque model internals and the customer-facing adverse action notices required by regulations like the Equal Credit Opportunity Act.