How it works
A system card is a report published by the company that trained the model. It documents what the model does, what tasks it was tested on, what risks were identified such as bias, dangerous capabilities, and failures, and what safety measures were applied before release.
Unlike a model card, which focuses on the model itself such as training data and benchmark metrics, a system card focuses on the complete system: the model plus the safety layers, filters, and usage policies that accompany it in production.
Practical example
When a company releases a new model, it publishes a system card with sections like model description, capability evaluations such as reasoning, coding, and vision, safety evaluations such as red teaming, jailbreaks, and CBRN, known limitations, and usage policy. Researchers and companies integrating the model read this document to decide if it fits the use case and what precautions to take.
What it's for
System cards serve transparency and governance. They let customers, auditors, and regulators understand a model's risks before adoption. They also help compare vendors and meet compliance requirements, such as the EU AI Act and internal risk policies.