AI 工程从零到一（中文）
Given a production deployment, recommend a moderation stack configuration across the three layers.
Produce:
- Input classifier. Choose OpenAI Moderation, Llama Guard 3/4, or Perspective API. Match to policy taxonomy. For multimodal deployments, Llama Guard 4 or OpenAI omni-moderation.
- Output classifier. Same or different from input classifier. Match thresholds to the downstream risk model.
- Custom domain rules. Enumerate the domain-specific rules the general classifiers will not catch: financial-advice disclaimers, medical-advice refusals, legal-disclaimer patterns.
- Judge for edge cases. Specify the human-escalation path. Hard refusals are final; ambiguous cases go to human review within SLA.
- Migration plan. If Azure Content Moderator is in the stack, plan the migration to Azure AI Content Safety before February 2027 retirement.
Hard rejects:
- Any deployment without output moderation (input alone is not sufficient).
- Any deployment without custom domain rules on regulated surfaces (finance, health, legal).
- Any deployment relying solely on pre-LLM-era classifiers (Perspective) for modern chat applications.
Refusal rules:
- If the user asks for the single best classifier, refuse — classifier choice is policy-taxonomy-specific.
- If the user asks for thresholds, refuse single numbers — thresholds depend on risk tolerance and downstream effect.
Output: a one-page recommendation filling the five sections, naming the classifier at each layer, and flagging migration obligations. Cite OpenAI Moderation docs and Llama Guard 3/4 references once each.