A model card is a structured document that travels with a machine learning model. It describes the model's intended use, the populations and conditions it was validated on, known limitations, fairness considerations, and the metrics used to evaluate it. A data card does the same thing for a dataset: provenance, collection conditions, representativeness, known biases, and terms of use. Together, they make a model or dataset legible to people who did not build it.
The practice emerged from a recognition that releasing a model or dataset without context is not neutral. It shifts the burden of interpretation onto the recipient, who may lack the context to assess risk, fitness for purpose, or potential for misuse. Structured documentation reduces that burden and makes accountability concrete rather than implicit.
What makes a card useful
A useful card is honest about limitations, not only strengths. It reports performance on subgroups, not only overall. It names the conditions under which the artifact should not be used. It is versioned so that updates are traceable. And it is legible to its intended audience, which may include clinical researchers, policy reviewers, or community advocates who are not machine learning specialists.
Writing a card that is honest about limitations requires institutional support. If an organization only rewards performance improvements and treats limitation documentation as a liability, its cards will be incomplete. Cytognosis treats comprehensive documentation as a condition of release rather than an optional supplement.
Open notebook
We are developing our own model card and data card templates adapted to the specific needs of health-state mapping research. This page is part of our open notebook and will be updated as templates are finalized and first cards are published.