DATA & MODEL PROVENANCE
Do you maintain model cards, and do they disclose known limitations?
Current as of July 2026.
A model answer
Yes. Each model in production has a model card recording intended use and out-of-scope use, the evaluations run and their results, performance across relevant groups, known limitations, and the date of the last review. Limitations are disclosed rather than omitted; a model card that claims no limitations is not a model card. Cards are versioned and re-issued when the model changes materially.
Evidence a buyer expects
The card itself, versioned, with a real limitations section. A reviewer reads the limitations section first — its absence or emptiness is the tell.
Why weak answers fail
Over-claiming — "no known limitations", "bias-free" — is read by a professional reviewer as either naivety or deceit, and both end the review. Honest limitations, paired with the controls that mitigate them, build far more confidence than a clean sheet nobody believes.
What Hael produces
Hael's Model Card (EU AI Act Article 13, NIST MEASURE): capability, limitations, group performance — generated from evaluation records and sealed.
Related questions
Current as of July 2026. General information, not legal advice.
Keep reading
More from the answer library.
OVERSIGHT & CONTROLS
How do you test for bias, and what did the testing find?
Read answer →
DATA & MODEL PROVENANCEDescribe the provenance of your training data and the rights under which it is used.
Read answer →
OVERSIGHT & CONTROLSWhere does human oversight sit in your AI system, and who holds it?
Read answer →