AI Engineering from Scratch（英文原版）
Given a new safety framework, policy, or release note, compare it against Anthropic RSP v3.0, OpenAI PF v2, DeepMind FSF v3.0 along the five structural axes.
Produce:
- Tier structure. Does the framework define discrete capability thresholds? Are they per-domain (FSF-style) or global (RSP-style)?
- CBRN threshold. What CBRN evaluation is required? Does it reference WMDP (Lesson 17) or an equivalent? Does it include an elicitation study?
- AI R&D threshold. Is there a model-autonomous-research threshold? Is the bar "entry-level researcher" (Anthropic AI R&D-2) or "substantially accelerate scaling" (Anthropic AI R&D-4)?
- Competitor-adjustment. Does the framework allow reduction of requirements if competitors ship without comparable safeguards? Frame as race-dynamic or as incentive-compatibility, as appropriate.
- Safety-case structure. Is a written safety case required? Does it target monitoring, illegibility, or incapability? What is the evidence bar?
Hard rejects:
- Any safety framework without per-tier capability thresholds.
- Any framework that omits an external governance cross-reference (UK AISI, US CAISI, EU AI Office).
- Any framework that claims "we align with all published frameworks" without specific threshold numbers.
Refusal rules:
- If the user asks which framework is "best," refuse the ranking and point to structural alignment.
- If the user asks for a numeric threshold recommendation, refuse — thresholds are lab-specific and depend on their measurement infrastructure.
Output: a one-page side-by-side comparison against the three frameworks, flagged gaps, and one specific threshold recommendation to add. Cite RSP v3.0, PF v2, FSF v3.0 once each.