02 / METHOD
How the translation
was made.
Source material, instructions, AI generation, reference alignment, and export—each with a different role and a different evidence boundary.
Nichijoyaku was created using AI. A complete review by human biblical-language and biblical-studies specialists has not been established. Read alongside established translations and original-language resources.
FROM SOURCE TO RELEASE
Follow the code’s approach
A map of the saved code. Open each stage to inspect its implementation. This does not establish that every run used every optional feature.
- 01
Source texts & references
Hebrew and Greek source material, with available reference translations.
prompt.py → - 02
Assemble the context
A verse or passage (pericope), surrounding context, glossary, and available translation memory.
prompt.py → - 03
- 04
Generate with Claude
Generate Japanese and process the response. Transient call failures have a retry path.
translator.py → - 05
Save & evaluate output
Store translation and metadata. The code also supports optional automated evaluation.
translator.py → - 06
Align & repair references
Map source references to reader-facing numbering; examine conflicts and missing passages.
alignment.py →Unresolved items?↳ Yes → repair and re-audit. Hold publication.↓ No → proceed to export - 07
Export & verify artifacts
Check the audit, book hashes, references and text before producing reader assets.
verify_aligned_release.py →
Passing hash and structure checks ≠ proof of semantic or theological accuracy
01 Source material and reference translations
The system documentation identifies MorphGNT/SBLGNT for the Greek New Testament and MorphHB for the Hebrew Old Testament. It lists KJV, ASV, Kougo, Meiji, and Taisho as reference translations. Availability to the system does not establish that every resource was supplied to every generation call.
The prompt builder supports surrounding context, a glossary, translation memory, and reference material. This collection cannot reconstruct the exact input to every call.
Supporting record ↗ translator-readme.md02 Policy and prompts
Instructions emphasize everyday vocabulary, natural sentences, genre-sensitive writing, and continuity across verse boundaries. Simplifying away the source’s content is not the stated aim.
Snapshots of the configuration and prompt-building code are available. They establish the saved code state, not that every historical generation used that exact configuration.
Supporting record ↗ prompt.py.txt03 AI generation and review
The translation system uses Claude. The alignment audit records reviewer labels including “claude-opus-5 via claude -p” and Codex local comparison/adjudication. These are tool labels in a record, not the names of human specialists.
Those labels do not establish the model used for every initial translation. A complete generation-model history, all prompts and responses, and total project expenditure are not supplied in this collection.
Supporting record ↗ alignment-summary.json04 Aligning references
Original-source numbering and reader-facing numbering can differ, including in Psalms and the prophetic books. Edition 1 is exported from aligned release artifacts, rather than directly publishing the raw source-numbered output.
The audit summary records 31,307 source records, 31,101 destination references, and 1,952 alignment changes. Its 206 extra references describe source-numbering differences; they should not be described as 206 deleted verses.
Supporting record ↗ alignment-summary.json05 Fixing and checking the artifacts
Each book JSON has a SHA-256 checksum. The reader’s export checks the manifest and audit hashes before generating its CSV. Preparing this evidence collection included checks of all 66 book hashes, book/chapter/verse counts, the audit hash, and the CSV hash.
Verifying saved artifacts is different from regenerating identical wording with a model. Full reproducibility of AI generation is not claimed.
Supporting record ↗ release-verification.jsonRecovered generation records
We publish 28,602 saved pre-alignment outputs from 1,096 chapters matching the release audit hashes. Of these, 6,446 retain prompts. Project-authored system instructions are available in full; user inputs have hashes and reference labels. Exact user prompts embed separately licensed Bible text and are not redistributed here.
Retained inputs include a Kougo label in 6,382 records and a KJV label in 6,390. These records contradict a claim that no reference translations were supplied. Ninety-four source chapter files are missing or changed and are excluded from release-linked outputs. Saved outputs are parsed translations, not raw model responses.
Supporting record ↗ retained-provenance.jsonBrowse the records by chapter ↗
Saved outputs and input hashes (JSONL) ↗