System Card¶

Purpose¶

A System Card documents the complete AI system — not just the model, but the full stack: model, retrieval configuration, tooling, safety layers, observability, and human oversight. It complements the Technical Model Card, which focuses on the model component alone.

Use this template when the system involves any of:

• Multiple models or model calls in a pipeline
• RAG (retrieval-augmented generation) with managed knowledge sources
• Tool use / function calling / external API access
• Agentic behaviour (Mode 3–5)
• Any High or Limited Risk classification

The System Card is a Gate 3 deliverable for High Risk systems and must be kept up to date through the Management & Optimisation phase.

Part 1: System Identity¶

System name: [Human-readable name, e.g. "Customer Claims Triage"] System version: [Semantic version, e.g. v2.1.0] Owner (Tech Lead): [Name] Guardian: [Name] Last updated: [YYYY-MM-DD] Risk level: [Minimal / Limited / High / Critical] Collaboration Mode: [Mode 1–5]

Part 2: Intended Use¶

Primary Purpose [REQUIRED]¶

[One paragraph describing what the system does, who uses it, and in which context.]

Intended Users¶

[Describe the primary user group(s) and their technical level.]

Out of Scope [REQUIRED]¶

List use cases this system must not be used for:

• [Prohibited use 1]
• [Prohibited use 2]

Hard Boundaries [REQUIRED]¶

Cross-reference to the Objective Card:

Part 3: System Architecture¶

Model Layer¶

Retrieval Layer (RAG)¶

Complete only if RAG is used.

Tool / Function Layer¶

Complete only if tool use or function calling is enabled.

Orchestration Layer¶

[Describe multi-step pipelines, agent loops, or chaining logic. Include a diagram if the flow is non-trivial.]

Part 4: Safety Layers¶

Input Guardrails [REQUIRED for High/Critical Risk]¶

Output Guardrails [REQUIRED for High/Critical Risk]¶

System Prompt Location and Access Control¶

• Location: [Version-controlled path / secret manager key]
• Access: [Who can read / modify the system prompt?]
• Change procedure: [Link to change management procedure]

Refusal Logging¶

• Refusals logged: [Yes / No]
• Log location: [System / retention period]

Part 5: Human Oversight¶

Oversight Mode [REQUIRED]¶

[Describe how humans oversee the system's outputs per the collaboration mode.]

Override Capability¶

• Override available to users: [Yes / No]
• Override mechanism: [How can a human correct or reject the output?]
• Override log: [Is the override logged? Retained how long?]

Escalation Path¶

[Document who is contacted when the system behaves unexpectedly.]

Part 6: Observability¶

Metrics Monitored¶

Log Retention¶

Part 7: Known Limitations¶

[REQUIRED] Document what the system cannot do reliably:

Part 8: Compliance Status¶

Part 9: Version History¶

Related Artefacts¶

• Objective Card
• Technical Model Card
• Validation Report
• Risk Pre-Scan
• Guardian Review Checklist