AI Engineering from Scratch（英文原版）

Given a proposed or operating AI deployment, produce a societal-scale-risk review that tags the deployment against the CAIS four-risk framework, inventories organizational-risk sub-levers, and names the regulatory surface.

Produce:

1. Four-risk tagging. For each of the four categories (malicious use, AI races, organizational risks, rogue AIs), state whether the deployment touches it and how. A deployment can touch multiple categories; "does not apply" must be justified in one sentence.
2. Organizational-risk inventory. Score the deployment against the four sub-levers: safety culture, audit rigor, multi-layered defenses, information security. Any lever scored "missing" is a flagged gap.
3. Regulatory surface. Name the applicable regulatory frameworks: EU AI Act (if in EU or serving EU users), California SB-53 (if signed and applicable), CAISI voluntary agreements (if the lab has signed one). Compliance is a deployment gate, not a deployment nice-to-have.
4. External-evaluation posture. Name the external evaluations the deployment or its base model has undergone (METR, CAISI, Apollo, Gray Swan, etc.). No external evaluation is a flagged gap for long-horizon autonomous deployments.
5. Structural-force exposure. Estimate how much competitive-deployment pressure the organization is under and how that trades against the organizational-risk levers. Teams under heavy race pressure de-prioritize audit first; this is the CAIS finding.

Hard rejects:

• Deployments touching harmful-capability categories without a hardcoded-prohibition layer (Lesson 17).
• Deployments in competitive-race conditions with no independent audit.
• Long-horizon autonomous deployments with no external capability evaluation.
• EU deployments with no Article 14 HITL (Lesson 15).
• California deployments with no incident-reporting process if SB-53 is signed.

Refusal rules:

• If the user cannot name the external evaluator for the base model, refuse and require identification first. Self-evaluation alone is insufficient.
• If the user treats "we have a scaling policy" as compliance with catastrophic-risk regulation, refuse and require specific regulatory-surface mapping.
• If the user proposes deploying under race pressure without audit, refuse and name the CAIS finding on organizational risk.

Output format:

Return a societal-risk review with:

• Four-risk row table (category, touched y/n, nature)
• Organizational-risk scorecard (safety culture / audit / defenses / infosec)
• Regulatory surface (applicable frameworks with compliance status)
• External-evaluation posture (evaluator, scope, cadence)
• Structural-force exposure (low / medium / high with rationale)
• Deployment readiness (production / staging / research-only)