/safety-case

name: safety-case
description: >-
  Assemble a structured assurance / safety case for deploying an AI system — an
  explicit argument that it is acceptably safe for its context, backed by evidence
  (harm model, evals, guardrails, fairness, governance). Use to support a
  go/no-go deployment decision or an audit/sign-off.

Goal

A defensible safety case: a top-level claim that the system is acceptably safe for its intended use and context, decomposed into arguments, each supported by concrete evidence and with residual risks stated honestly.

Structure (claims → arguments → evidence)

1. **Top claim** — "System X is acceptably safe to deploy for use case Y in context Z," with the acceptance criteria and who owns the decision. 2. **Sub-claims / arguments** — typically:

• Harms are identified and bounded (evidence: `ai-safety:harm-modeling`).
• Safety behavior is measured and meets thresholds (evidence:

`ai-safety:safety-evaluation`, `ai-safety:safety-red-team`).

• Treatment is equitable (evidence: `ai-safety:bias-fairness-assessment`).
• Guardrails and human oversight are adequate (evidence:

`ai-safety:guardrail-review`).

• Governance, monitoring, and incident response are in place (evidence:

`ai-safety:responsible-ai-assessment`). 3. **Evidence** — link each argument to the actual artifacts/results; note their date, scope, and limitations. 4. **Residual risk & conditions** — what remains, why it's acceptable, and the conditions/monitoring/kill-switches that keep it acceptable post-deployment.

Steps

1. State the deployment context and acceptance criteria first (they set the bar). 2. Build the argument tree; for each leaf, attach evidence or mark it as a gap. 3. Make gaps explicit — a safety case with honest gaps beats one that hides them. 4. Conclude with a clear go / no-go / go-with-conditions recommendation.

Output

A safety-case document (claims → arguments → evidence → residual risk → recommendation). Render the argument tree with `security-diagramming` and produce the document + executive summary with `security-reporting`.

Notes

A safety case is an honest argument, not a checklist or a rubber stamp. Unsupported arguments and unaddressed residual risks are the point — surface them so the decision-maker owns them explicitly.