Skip to content

Latest commit

 

History

History
44 lines (31 loc) · 2.17 KB

File metadata and controls

44 lines (31 loc) · 2.17 KB

Glossary Audit Workflow

AcademyLens glossary packs are useful only when reviewers can find mistakes quickly. The repo includes an audit packet generator for AI review, native review, and community review.

Generate Audit Packets

npm run glossary:audit
npm run glossary:audit -- --locale=ja

The command writes Markdown and JSON packets to dist/glossary-audit/. The folder is ignored by git.

Each packet includes:

  • every source/target pair for the locale
  • source IDs used for the term
  • review flags such as single-word-overreach-risk, latin-target-review, and target-same-as-source
  • protected-term collision checks

Review Priorities

  1. Check single-word-overreach-risk terms first. Short words such as model, module, lesson, and examples can over-apply in normal course prose.
  2. Check terms with Latin text in the target language. Some are intentional, such as JSON mode, but others may be lazy transliteration.
  3. Compare high-impact terms against OpenAI Academy pages and OpenAI developer documentation.
  4. For llm-drafted packs, mark issues without promoting the pack status.
  5. For llm-audited packs, confirm the second-pass audit signal and run npm run check:glossary-quality.
  6. Promote a pack only after native/community review or a documented review process.

Status Meanings

  • llm-drafted: AI-assisted draft. Good enough for beta correction, not enough to claim native quality.
  • llm-audited: A second AI review pass has been applied, recorded, and locked by npm run check:glossary-quality.
  • community-reviewed: Reviewed by a contributor with language knowledge.
  • native-reviewed: Reviewed by a native or near-native speaker.
  • reviewed: Final project-maintainer status after review evidence is complete; community-reviewed packs remain below this level until pending QA signals are closed.

Over-Translation Smoke Check

npm run check:glossary-overreach

This script runs selected Academy/Gradual fixture text through the glossary masker and fails if glossary placeholders become suspiciously dense. It is not a substitute for manual DOM QA, but it catches the most obvious "the dictionary got too aggressive" regressions.