AI 工程从零到一（中文）

Given a specific frontier capability (e.g., "long-range autonomy," "autonomous replication and adaptation," "R&D automation"), produce a cross-policy diff showing how each of the three frameworks classifies the capability and what mitigations trigger.

Produce:

1. OpenAI PF v2 classification. Tracked or Research. If Tracked, name the Capabilities + Safeguards Report triggers. If Research, note the policy language is "potential" mitigations.
2. Anthropic RSP v3.0 classification. Which threshold (ASL-3, AI R&D-4, hardcoded prohibition)? Which mitigation (affirmative case, security + deployment)? Confirm whether the commitment lives in the Anthropic-unilateral tier or the industry-recommendation tier.
3. DeepMind FSF v3 classification. Which domain (Cyber, Bio, ML R&D, CBRN)? Which CCL or Tracked Capability Level? Is deceptive alignment monitoring invoked?
4. Convergence summary. Do the three policies agree on the capability's severity, or is there meaningful disagreement? Which classification is most rigorous, which least?
5. Measurement dependency. Every classification depends on capability measurement. Name how the capability is measured and which eval provider (METR, Apollo, internal, third-party) owns that measurement.

Hard rejects:

• Claims of cross-policy alignment based on announcement-language similarity without document-level evidence.
• Any classification that cannot point to a specific clause in the source document.
• Treating "Research Category" (OpenAI) as equivalent to "Tracked Category" — they have different operational consequences.

Refusal rules:

• If the user cannot produce the source document passages for each classification, refuse and require citations first.
• If the user treats policy-existence as evidence of mitigation-in-practice, refuse and require evidence of the specific mitigations firing.
• If the capability is claimed to be "covered" by a framework but the word does not appear in the document, refuse and require a concrete clause reference.

Output format:

Return a diff document with:

• Capability definition (one sentence)
• OpenAI PF v2 row (classification, trigger, source clause)
• Anthropic RSP v3.0 row (classification, trigger, unilateral-vs-recommendation)
• DeepMind FSF v3 row (domain, CCL / TCL, deceptive-alignment involvement)
• Convergence summary (agreement + meaningful disagreement)
• Measurement ownership (eval provider, eval cadence)
• Reader recommendation (most rigorous, least rigorous, justified)