Model Card
A short document reporting what a model is for, how it was evaluated, and where it performs worse. An industry convention, not an NAIC requirement.
A model card is a short document that travels with a trained model. The idea comes from a 2019 paper by Mitchell and co-authors, which proposed cards that give benchmarked evaluation across a variety of conditions, state the context the model is intended for, describe how it was evaluated, and report performance broken out by group rather than as a single headline number. The point of the format is that it makes the limits of a model as easy to find as its accuracy.
In an insurance governance program the card is the data science team’s contribution to the evidence file. It typically records what the system does, what it was trained on, how it was tested and against what, the known weaker segments, and the intended and out-of-scope uses. It is not, however, a substitute for sign-off: a card written by the team that built the model stays internal work product until someone with authority to accept the model puts their name to it, which is the boundary model validation exists to mark.
No NAIC document requires a model card by name, and the regulator asks instead for the underlying facts. The questions in the evaluation tool run to things like the system name, the type of AI, when it was implemented, whether it was built internally or bought, and when it was last tested. A model card is a convenient place to keep those answers so they exist before anyone asks, but the format earns nothing on its own. Keeping the card current after the model changes is what separates it from documentation theatre, which is the same discipline model validation and the model inventory rely on.