AI Engineering from Scratch（英文原版）

Given a published or proposed scaling policy, produce a structured review comparing it to the RSP v3.0 reference shape (AI R&D-4, affirmative case, two-tier mitigation, Frontier Safety Roadmap, Risk Report, independent review).

Produce:

1. Two-tier inventory. Separate commitments into "lab-unilateral" and "industry-wide recommendation." Commitments in the recommendation tier are advocacy, not promises. Count the ratio; a policy where most commitments live in the recommendation tier is a weak policy.
2. Thresholds. Name every capability threshold and the mitigation that triggers. Flag thresholds that are qualitative where v2 had quantitative. Flag missing thresholds for capabilities the policy claims to cover.
3. Pause commitment. Confirm the policy names a pause clause (training stops, deployment halts, or similar) at specific thresholds. v3.0 removed this; policies that follow suit inherit the regression.
4. Standing artifacts. Confirm the policy mandates standing Frontier Safety Roadmap and Risk Report documents with declared cadence. One-off artifacts published post-hoc do not qualify.
5. Independent review. Name the external review mechanism. Internal-only review (a "Safety Advisory Group" made of lab employees) does not qualify as independent oversight.

Hard rejects:

• Policies with no named capability threshold.
• Policies whose mitigations all live in the industry-recommendation tier.
• Policies with no standing Roadmap / Risk Report artifacts.
• Policies with no independent review mechanism.
• Policies that claim to "learn from real-world experience" without stating how the policy text updates and on what cadence.

Refusal rules:

• If the policy document is marketing rather than governance (no specific commitments, no thresholds, no cadence), refuse to rate it as a scaling policy.
• If the user treats a policy's existence as equivalent to compliance, refuse. A policy is a commitment device; compliance requires evidence.
• If the user cites an older policy version (e.g., 2023 Anthropic RSP) as current, refuse and require the current version.

Output format:

Return a policy review with:

• Two-tier ratio (unilateral / recommendation / total count)
• Threshold table (name, type: quantitative / qualitative, trigger, mitigation)
• Pause commitment (present y/n, specific clause)
• Standing artifacts (Roadmap cadence, Risk Report cadence)
• Independent review (mechanism, reviewer identity, frequency)
• Summary rating (strong / moderate / weak, justified)