Sacred text translation pipeline with human governance
Challenge: A team of 27 volunteer translators works to translate Baha'i sacred writings into Chinese. Working manually, they complete approximately 1 book per year. Each translation must meet three rigorous standards: accuracy (准确) — faithful to the original meaning; beauty (文风优美) — elevated, literary Chinese register worthy of sacred scripture; and consistency (风格一致) — uniform terminology and style aligned with the tradition established by Shoghi Effendi.
Solution: The Bahai Chinese Translation Workbench is an AI-assisted pipeline that accelerates translation throughput to an estimated 4-6 books per year while preserving human governance over every translation decision. The system implements a 3-stage pipeline — AI Translation, Human Review, and AI Editing — where no translation advances without explicit human approval.
Stage 1 generates a Chinese translation draft using Claude Sonnet 4.6. The prompt includes the full 20-term terminology glossary to ensure proper nouns, theological terms, and scripture titles are translated consistently.
Source text and Chinese translation are displayed in a two-panel layout. Reviewers can read the original and translation simultaneously, making it easy to verify accuracy and completeness.
Stage 2 places the translation under human governance. Reviewers can approve (pass as-is), edit (modify and pass), or reject (block and flag). No translation advances without a human decision.
Stage 3 refines the approved translation for grammar, punctuation, tone, and terminology uniformity. The AI editor evaluates against all three standards and reports a checklist with assessment notes.
A searchable sidebar displays all 20 approved glossary terms with English, Chinese translations, and usage notes. The glossary is accessible at any stage for quick reference.
Both the translation and editing prompts encode the three quality standards — accuracy (准确), beauty (文风优美), and consistency (风格一致) — directly into the system prompt. The AI evaluates and reports against each standard.
The 20-term glossary is formatted and injected into every AI prompt. Terms cover persons (Baha'u'llah, Abdu'l-Baha), institutions (Universal House of Justice), theology (Manifestation of God), and scripture titles (Kitab-i-Aqdas).
AI agents return structured JSON with translation text, term usage reports, and notes. A recursive unwrapping function handles cases where the LLM wraps output in markdown fences or nested JSON objects.
The backend enforces strict stage sequencing. Documents must complete Stage 1 before review, and Stage 2 before editing. The review endpoint validates the decision (approve/edit/reject) and routes accordingly.