# 日常訳 / Nichijoyaku

An open Japanese Bible translation, maintained and stewarded by **Conor Livingston**.

The Japanese text is **CC BY 4.0**; project software is **MIT**. See LICENSE
and software/LICENSE. The Bible was generated with AI and has not received
comprehensive human expert review. File verification does not prove accuracy.

## Read, reuse, contribute

- [Read edition 1](https://japanbible.net/?translation=nichijo-v1).
- [Project, method and evidence](https://nichijoyaku.pages.dev/).
- `edition-1/`: verified book JSON, JSONL, CSV, text, USFM, furigana and
  printable academic study layout. The study layout uses the same edition.
- `edition-1/checksums.json`: SHA-256 for every distribution file.
- Download the complete source archive from the project website; no GitHub
  account is needed to inspect, verify or reuse it.
- [Suggest a correction without an account](https://nichijoyaku.pages.dev/contribute/).
  Identify edition, passage, current wording or reading, proposed change and reason.
  Do not include private data.

## What can be inspected

`software/` contains the translation pipeline's source, not its private runtime
environment. `evidence/` includes pinned prompts/configuration, release evidence,
usage summaries, and surviving numbering decisions with Japanese alternatives.
`provenance-v1/` publishes 28,602 saved parsed outputs from 1,096 release-matched
chapter files. 6,446 include retained-input hashes, reference labels and timing;
one project-authored system prompt is available in full. Historical prompts
include Kougo and KJV, so a claim of no reference inputs is contradicted.
Exact user prompts remain private because they embed separately licensed text.
94 missing/changed chapter inputs are explicitly listed. Two Joel chapters
were recovered from Git history and verified against the original audit.
Original-language corpora and reference Bible translations are not redistributed.
Download prerequisites using the software's documented source importers.

The prompt builder supports reference translations. We do **not** claim that
no other translations were input, that all model training data was copyright
free, or that every historical LLM input/output survives. Existing evidence is
published as evidence; missing records are not reconstructed and relabeled as
original logs. Model labels in saved records are not independent vendor checks.

## Reproduce and verify

From the extracted source directory: `python3 tools/verify-package.py` checks every distributed
file against its checksum manifest. No network or model calls are required.
The release-generation tools in `tools/` document the export from the source
monorepo; they need that original layout and are not standalone translators.
Use `software/README.md` for pipeline requirements. Private runtime state is not
included and running a translation may incur provider charges.

## Changes and stewardship

Submissions do not rewrite a published edition automatically. A change records
the old text, proposed text, reference, reason, reviewer role and decision.
Keep edition 1 immutable; publish accepted translation changes as a new edition
with its own manifest. AI review and human review must remain distinguishable.

Why open source? Reuse should not depend on one website, device or maintainer.
People should be able to inspect the work, correct mistakes and build new
reading, teaching and accessibility tools around it.