AI 工程从零到一（中文）

Given a deployment's constitutional layer (system prompt, operator config, declared principles), audit it against the Claude Constitution reference and flag missing hardcoded prohibitions, ambiguous principles, or misordered tiers.

Produce:

1. Hardcoded prohibition inventory. List every prohibition that must not bend regardless of operator or user instruction. Minimum floor: bioweapons / CBRN uplift, CSAM, critical infrastructure attack planning, false-identity-when-asked. Additions are deployment-specific (e.g., financial services adds specific fraud prohibitions).
2. Soft-coded defaults. List every behaviour the operator can adjust. For each, state the declared bound. An "adjustable" setting with no bound is a back-door override.
3. Tier ordering. Confirm the resolution order is: safety > ethics > guidelines > helpfulness. If helpfulness ever wins over ethics in the implemented resolver, flag as a deployment break.
4. Principle ambiguity flags. Identify any principle whose text leaves room for materially different interpretations. Ambiguity compounds over training cycles (principle drift).
5. Layer completeness. Confirm runtime-layer controls (Lessons 10, 13, 14) are present in addition to the constitutional layer. Constitution alone is insufficient; runtime alone is insufficient.

Hard rejects:

• Deployments without any hardcoded prohibition layer.
• Operator config that claims to override a hardcoded prohibition (even by renaming).
• Tier orders that place helpfulness above ethics.
• Principle text so general it cannot be evaluated ("be good").
• Treating Constitutional AI as a replacement for runtime controls.

Refusal rules:

• If the user names a hardcoded prohibition but cannot point to a runtime-layer backstop for it, flag the deployment as single-layer and refuse production.
• If the operator config includes an adjustable "safety" setting with no declared bound, refuse.
• If the user treats the 2023 participatory-constitution findings as actionable in the current deployment, check: the 2026 Constitution did not incorporate them, so "inherits democratically" is a claim the deployment cannot back up.

Output format:

Return a constitutional audit with:

• Hardcoded floor (prohibitions, enforcement layer: weights / inference / both)
• Soft-coded defaults (setting, operator bound, user-visible y/n)
• Tier order (listed; confirmed safety > ethics > guidelines > helpfulness)
• Ambiguity flags (principle, specific ambiguity, proposed tightening)
• Layer completeness (constitutional y/n, runtime controls y/n, both required)
• Readiness (production / staging / research-only)