User Manuals
Select the manual for your role:
How We Translate
BibleCollab's translation happens in two distinct phases. In Phase 1, the AI pipeline generated a v0.1 draft for every passage (complete). In Phase 2 (current), community translators and editors review, discuss, and revise the v0.1 drafts, with AI serving as a reference aid while people make all editing decisions.
Phase 1: AI Generates v0.1 (Complete)
The following five steps were automated by the AI pipeline — no manual editing of the v0.1 draft text. v0.1 drafts for all 1,646 passages in both Chinese and English have been generated.
Data Collection
We gather source texts (Greek SBLGNT, Hebrew BHSA), reference translations (CUV, BSB, WEB), cross-references (TSK), lexical cards, historical-cultural cards, discourse structure cards, and commentary summaries to provide a complete data foundation for exegesis and translation.
Exegesis Agent
GPT-5.2 and Gemini 3 Pro Preview each independently consume all reference data and compile structured Exegesis IR (Intermediate Representation) for each passage, containing MUST/SHOULD/MAY rules covering lexical, grammatical, textual, and historical-cultural dimensions.
IR Merge
A pure Python pipeline merges the independently generated Exegesis IRs into a single authoritative Canonical IR: deduplicating identical rules, resolving conflicts via an Authority Ladder, and harmonizing severity levels. Each passage averages 42.8 rules.
Translation Agent
Multiple models consume the merged Exegesis IR and source texts to independently generate translations. The Chinese version is produced by GPT-5.4, Gemini 3.1 Pro Preview, and Doubao Seed 2.0 Pro, each revising from the CUV (Union Version) base. The English version is produced by GPT-5.4 and Gemini 3.1 Pro Preview, translating directly from the original languages.
Reviewer Agent
An AI reviewer anonymizes and blindly evaluates each model's translation, scoring faithfulness, fluency, and style to select the best rendition as v0.1. Chinese review is done by DeepSeek v3.2 (picking 1 of 3); English review by Claude Opus 4.6 (picking 1 of 2).
Phase 2: Community Collaboration (Current)
After v0.1 is published, the process becomes human-driven. Editors and community members review translations, discuss improvements, and submit revisions. AI no longer decides the final wording — instead it serves as a reference aid: you can browse the materials used to generate v0.1 (source texts, exegesis rules, discourse structures) and request AI commentary on discussions. All editing decisions are made by people. Every edit is version-tracked, and translations improve through iterative release cycles.
Community Review & Revision
Editors and contributors review v0.1 translations on the pericope page. Reference panels provide CUV/BSB comparisons, Greek interlinear analysis, exegesis rules, and discourse structures. Logged-in users can comment with suggestions; editors decide what to adopt and submit changes. Every submission automatically creates a new version.
Version Iteration & Release
Translations advance through discussion cycles: during a dev cycle, everyone discusses and editors refine the text. When ready, the admin publishes an official release (e.g., v1.0) — a frozen, immutable snapshot. A new dev cycle then opens for continued improvements. The entire workflow is similar to Git version control.