[PROPOSAL] Evaluate Traditional Chinese content with structured outputs

[ting-hong-shieh]

Preflight checklist

• I searched existing cookbooks and issues; this topic is not already covered.
• This proposal focuses on Claude API/SDK capabilities, not an external framework.
• I am prepared to implement the cookbook.

Problem statement

Teams producing Traditional Chinese content need a review process that separates locale errors and unsupported wording from subjective style preferences. Phrase blacklists tend to overcorrect: a phrase may be appropriate in formal copy, and punctuation frequency is not evidence that text was AI-generated.

This cookbook would show how to build an evidence-based review loop for zh-Hant-TW content. Each finding must identify the affected span, name the applicable rule, and propose the smallest revision that fixes it.

Claude capabilities demonstrated

• Structured outputs with client.messages.parse() and a Pydantic result model
• Rubric-guided evaluation with quoted evidence for each finding
• A second Messages API call that revises only supported findings
• Batch evaluation of a small, local fixture set
• Separation of deterministic locale checks from model judgment

Learning objectives

Readers will learn how to:

1. Define an evaluation schema that makes findings inspectable.
2. Write locale-specific criteria without treating a phrase list as an authorship detector.
3. Preserve facts, product names, commands, and deliberate voice during revision.
4. Include counterexamples so formal or technical prose is not penalized automatically.
5. Compare the original and revised copy in a compact scorecard.

Proposed outline

1. Scope and limitations
2. A small zh-Hant-TW fixture set with both defects and counterexamples
3. Deterministic checks for locale terminology and punctuation consistency
4. A structured-output evaluator
5. Evidence-gated revision
6. Reviewing false positives and adjusting the rubric
7. Cost and reproducibility notes

The notebook would be self-contained and use only the Anthropic Python SDK, Pydantic, and pandas. The live example would use a small number of short API calls and state the expected cost before execution.

Differentiation

The repository contains evaluation and classification examples, but I could not find a cookbook focused on Traditional Chinese content quality or on separating locale checks from subjective editorial judgment. This proposal also avoids framing stylistic signals as proof of AI authorship.

Implementation plan

I am prepared to implement the notebook, fixture data, registry entry, and author metadata, then run the repository's notebook validators.

[ting-hong-shieh]

An implementation is available in draft PR #827.

Local validation completed successfully for notebook execution without an API key, Ruff linting and formatting, notebook structure, author ordering, and registry schema and paths. The live review and revision branches remain untested because the local environment did not contain ANTHROPIC_API_KEY.

I am keeping the PR in draft while this proposal is reviewed. I can adjust the scope or examples before marking it ready.