Model Card

Basic metadata about the model artifact, including its version, type (e.g., convolutional neural network, transformer), and the organization or team that developed it. This section also specifies the date of training and the hardware/software stack used, providing a unique identifier for reproducibility and lineage tracking. It answers the fundamental question: Who built what, and when?

A precise specification of the use cases the model was designed and validated for, including the target domain and user base. This section explicitly states out-of-scope applications where the model's behavior is undefined or likely to fail, such as using a consumer sentiment classifier on clinical text. It establishes the operational envelope for safe deployment.

An inventory of relevant demographic, phenotypic, or environmental factors that influence model performance. This includes instrumentation variables (e.g., camera type, sensor calibration) and population subgroups (e.g., age, dialect, signal-to-noise ratio). The section identifies which factors were analyzed for disparate impact and which were not, making evaluation scope transparent.

Quantitative performance results presented using real numbers and confidence intervals, not just aggregate scores. Metrics should be disaggregated by the factors listed above, showing performance across subgroups. Common metrics include:

• Accuracy, Precision, Recall, F1 for classification
• Mean Absolute Error (MAE) for regression
• Equal Opportunity Difference for fairness This section reveals where the model excels and where it breaks.

A detailed description of the datasets used to produce the reported metrics, including their source, collection methodology, and temporal coverage. This section distinguishes between training data, validation data, and held-out test data, and notes any known distributional mismatches between evaluation data and the intended deployment environment, such as covariate shift or concept drift.

A candid discussion of known limitations, failure modes, and potential negative impacts. This includes sensitivity analysis to adversarial perturbations, uncertainty quantification gaps, and risks of selection bias or confounding bias in the training data. The section documents any mitigation strategies implemented and identifies unresolved risks requiring human oversight.