System Card

Unlike a Model Card, which evaluates a model in isolation, the System Card documents the safety performance of the integrated system. This includes the model combined with its guardrail configurations, output moderation APIs, and human-on-the-loop oversight mechanisms. It reports on end-to-end testing where adversarial inputs test the entire pipeline, not just the raw model inference. Key metrics include the jailbreak susceptibility of the wrapped system and the grounding score when retrieval-augmented generation is active.

This section defines the precise intended use and out-of-scope applications for the deployed system. It specifies:

• Deployment environment: Cloud API, air-gapped on-premise, or edge device
• User personas: Who is authorized to interact with the system
• Input modalities: Text, image, audio, or multimodal combinations
• Rate limits and concurrency: Operational throttling to prevent abuse
• Kill switch mechanism: The hard-coded shutdown protocol for critical failures This context is essential for conformity assessment under the EU AI Act.

A System Card includes a detailed AI Bill of Materials (AIBOM) listing every component in the algorithmic supply chain:

• Foundation model: Version, provider, and model provenance
• Guardrails: Content filters, toxicity classifiers, and PII redactors
• Retrieval pipelines: Vector databases and knowledge graphs used for grounding
• Tool integrations: APIs and functions the system is authorized to call
• Human oversight interfaces: Dashboards for human-on-the-loop intervention This inventory enables vendor due diligence and third-party auditing.

This component discloses the results of structured red-teaming exercises conducted on the complete system. It documents:

• Prompt injection vulnerability testing across all input channels
• Indirect injection attacks via retrieved documents or tool outputs
• Multimodal jailbreak attempts combining image and text vectors
• Alignment faking detection protocols used during evaluation The report quantifies the system's adversarial robustness benchmark scores and describes any specification gaming behaviors observed during testing.

The System Card reports on disparate impact ratios and other fairness metrics measured at the system output level. Unlike model-level evaluations, this accounts for how guardrail configurations and output moderation APIs may introduce or mitigate bias. It includes:

• Demographic parity across protected groups
• Equalized odds for high-stakes decision flows
• Counterfactual fairness testing results
• Concept drift monitoring thresholds that trigger re-evaluation This section directly supports algorithmic impact assessments required by procurement policies.

This section defines the post-market surveillance plan and operational resilience protocols. It documents:

• Data drift detection pipelines and alerting thresholds
• Hallucination rate benchmarks monitored in production telemetry
• Rollback procedure: Steps to revert to a previous stable system version
• Model deprecation policy: Timelines and migration support commitments
• Incident response escalation paths and containment strategies The System Card serves as a living document, updated when any component in the algorithmic supply chain changes.