Data
Every model we deploy for an agency ships with a model card
A short, plain-language document that explains what a model does, how it was tested, and where it should not be used.
A model card is a one- or two-page document that accompanies a machine learning model. It states what the model is for, what data it was trained on, how it performs, and what its limitations are. We treat it as a required deliverable for any model we build for a public agency.
Why agencies need it
Automated decisions that affect residents are subject to public accountability, and increasingly to statute. When a council member, a journalist, or a resident asks how a model works, the answer should already be written down in language they can read.
What ours contain
- Purpose. The specific decision or task the model supports, and the human role that remains responsible.
- Data. Sources, time period, known gaps, and how sensitive attributes were handled.
- Performance. Accuracy and error rates overall and across relevant groups, with the test methodology.
- Limitations. Situations where the model should not be relied on, and known failure modes.
- Monitoring. What is measured in production and what triggers retraining or shutdown.
- Appeal. How a person affected by a model-supported decision can seek human review.
How it changes the work
Writing the card forces questions early: is this a problem a model should solve, what does fair mean here, what happens when it is wrong? Projects that cannot produce a defensible card should not proceed, and it is better to learn that in week two than at launch.
Alignment
The practice aligns with the NIST AI Risk Management Framework and with the direction of state guidance on public-sector AI. It is also simply good engineering.