AI 工程从零到一（中文）

Given a use case (faithfulness check, zero-shot classification, document-level inference), output:

1. Model. Named NLI checkpoint. Reason tied to domain, length, language.
2. Template (if zero-shot). Verbalization pattern. Example.
3. Threshold. Entailment cutoff for the decision rule. Reason based on calibration.
4. Evaluation. Accuracy on held-out labeled set, hypothesis-only baseline, adversarial subset.

Refuse to ship zero-shot classification without a 100-example labeled sanity check. Refuse to use a sentence-level NLI model on document-length premises. Flag any claim that NLI solves hallucination — it reduces it; it does not eliminate it.