Best Practices
- 1Always specify which locales and which atom banks before starting a wave
- 2Verify coverage percentages against the real atom count — never trust reported numbers without a grep
- 3For zh-TW, use Claude only — Qwen emits Simplified Chinese and is banned for zh-TW
- 4CJK quality gates are locale-aware — English scorers block CJK production incorrectly
- 5Track stubs separately — a translated stub is not a translated atom
CLI Examples
# Run a translation wave act as pearl_localization — run wave for zh-TW, ja-JP covering SOURCE_OF_TRUTH/social_media_atoms/evergreen_en_us_atoms.jsonl # Check coverage pearl_localization: what is the real zh-CN coverage percentage (grep, not reported)? # Dispatch a locale agent pearl_localization: dispatch Pearl_Locale_JA for atoms 1000-2000 in the evergreen bank # Quality gate pearl_localization: run quality gate on zh-TW batch — check Big5 encoding and no Simplified leakage
Do Not Use For
- Individual atom translation — dispatch to locale subagents
- Treating Qwen output as zh-TW — always verify Big5 encoding
- Reporting 100% coverage without byte-counting the actual translated rows