Resources · Worked example 4
Model card
A model card is easy to fake by leaving things out. This one is built to do the opposite: the credibility comes from the findings a vendor brochure would hide.
PriceWise prices term life and health policies, which makes it high-risk, and Vesta is both its builder and its user. The card shows the amber items plainly. The youngest age band is thin in the training data and slightly under-priced, so its manual-review band is widened rather than tightened. Rural-county residents sit borderline on calibration and are flagged for the next retrain. Gender is excluded as an input, following Test-Achats, and a proxy-leakage test checks whether occupation and disclosed conditions reconstruct it anyway: the reconstruction score of 0.64 sits under the 0.70 action threshold and is recorded so the next test has a baseline. None of that is hidden in an appendix; it is the point of the document.
The oversight is specific, not aspirational. Underwriters see the top contributing factors with every score and can override with a recorded reason; the override rate is reviewed monthly, and a rate above 15% in any segment pauses the model there. A monthly drift dashboard, quarterly fairness tests, and an annual revalidation feed board reporting, and every retrained version gets a new card before it ships. The structure follows the Mitchell et al. model-card format, adapted for an insurer that is both provider and deployer under the AI Act.
Legal references: EU AI Act Annex III 5(c) (insurance pricing) · Article 27 (fundamental rights impact assessment). Card format: Mitchell et al., Model Cards for Model Reporting (2019).
What's inside
- Model details, architecture, and regulatory status (provider and deployer)
- Intended use, and the out-of-scope uses drawn explicitly
- Training data, the excluded inputs, and the known data gaps
- Performance by segment, with Gini and calibration explained
- Fairness evaluation: proxy-leakage, distribution, and override-pattern tests
- Limitations, drift monitoring, and the amber findings carried openly
- Human oversight, monitoring cadence, and card maintenance
The other four examples
Need this documented for your own company before August 2?
That's what the literacy and governance workshop produces: your own governance documents and the training record behind them.