A system card documents an AI application as deployed, not just the model, but the whole system: architecture, data flows, guardrails, evaluation results, human
Documentation for an AI application as deployed, not just the model, but the whole system: architecture, data flows, guardrails, evaluation results, human-oversight design, and residual risks. It answers what the system does in production and how it can fail.
A model card documents a model; a system card documents the deployed application. System cards became the documentation unit regulators and buyers actually want, since real-world risk lives in the application layer, tools, data, autonomy, more than in raw model weights.