Write a short card for every AI feature

How to do it

The success spec in define success before you build says what the feature must do. The card says what it does now. Support, legal and the next developer read it when a user complains or a provider changes a model. Fill the scores from the latest eval report in CI, not by hand. List each model with its retirement date from track the retirement date of every model you call. Models you train yourself get the full model card.

Examples

TypeScript example

const CARD_SECTIONS = ['Models', 'Data read', 'Intended use', 'Out of scope', 'Eval scores', 'Known limits', 'Owner'];

export function missingCardSections(cardMarkdown: string): string[] {
  return CARD_SECTIONS.filter(section => !cardMarkdown.includes(`## ${section}`));
}

Python example

CARD_SECTIONS = ["Models", "Data read", "Intended use", "Out of scope", "Eval scores", "Known limits", "Owner"]

def missing_card_sections(card_markdown: str) -> list[str]:
    return [section for section in CARD_SECTIONS if f"## {section}" not in card_markdown]

How to check

A CI script finds a card for every feature in the gateway's prompt list. It fails when a card lacks any of the seven sections, or when its scores differ from the latest eval report.

Sources

• NIST AI 100-1, AI Risk Management Framework

Related rules

• AIF-2 Define success before you build in AI features
• AIF-75 Track the retirement date of every model you call in AI features
• MLO-22 Write a model card for every model in production in Machine learning