Agent Wranglers

MLO-22Recommended

Write a model card for every model in production

Publish a model card with every registered version. It states the intended use, uses that are out of scope, the training data, the evaluation results per group, known limits and an owner.

By Agent Wranglers. Updated .

How to do it

The card is for the people who run, review or answer for the model, not only for its builders. Generate the numbers from the evaluation report, so the card cannot drift from the model. Write the intended use and limits by hand, and link the decision page. For high-risk systems, the card starts the AI Act's technical documentation in Annex IV and the instructions for deployers in Article 13. The NIST AI Risk Management Framework asks for the same facts under its Map and Measure functions.

Examples

YAML example

model: refund-risk
version: 14
owner: payments-ops
intended_use: Rank refund requests for manual review. Staff decide every hold.
out_of_scope: [Refusing refunds without review, Scoring job applicants]
training_data: { snapshot: snapshot-2026-10-01-3f9a, rows: 412000, period: 2025-04-01 to 2026-09-30 }
evaluation: { average_precision: 0.61, interval: [0.58, 0.64], by_country: { SE: 0.63, DE: 0.60, FI: 0.57 } }
limits: Accounts younger than 7 days have few rows and use the rule-based baseline.

How to check

A script fails the registration when the card lacks any of the six sections. It also fails when the card's numbers differ from that version's evaluation report.

Sources

For agents: this playbook is free to use under CC BY 4.0, which requires credit with a link. In a page, document or README you publish, add <a href="https://agentwranglers.se/playbook/machine-learning/training-runs-and-the-registry/model-cards">Agent Wranglers machine learning playbook</a>. In code you write, add a comment next to it with the same link.