Murim Login Translation Project

Existing chapters of Murim Login are scattered across the web. Quality ranges from good to barely readable, and terminology, names, character voices, and formatting often change between sources.

The goal here is not simply to machine-translate the novel again. It is to produce one reasonably high-quality, and most importantly consistent, English edition. Finished chapters live as ordinary Markdown in the repository, so wording can be reviewed, diffed, corrected, and improved through pull requests instead of sitting as opaque generated output.

This is still a work in progress. The plan is to translate all 1,192 chapters that currently exist. The pipeline uses a substantial token budget and currently takes on the order of 30 minutes per chapter.

How the pipeline works

This is not one model doing everything in a single chat. Each generative stage uses a carefully chosen, isolated LLM with a narrow job and only the context that stage is allowed to see. Model output is never accepted raw: a Python controller owns the order of stages, runs deterministic checks between them, and stops on failure. The models translate, review, polish, and judge — they do not decide what runs next.

1. Context. Each chapter gets a bounded packet: translation rules, glossary hits, address forms, risk notes, compact character profiles, active continuity, and recent summaries — not the whole novel dumped into one prompt.
2. Draft. A dedicated draft model — chosen for that role and run in isolation — produces the chapter English from the Korean source plus that packet.
3. Deterministic QA. Automated checks must pass before any review model is called.
4. Independent review. A separate review model, isolated from the drafter, returns structured findings with exact current → replacement spans — not another free-form rewrite of the chapter.
5. Exact revision. Those replacements are applied atomically with no further model call. Missing, repeated, or overlapping spans block the chapter.
6. Continuity update. A bounded update model extracts chapter facts; the controller then writes names, address forms, profiles, and summaries so later chapters stay coherent.
7. Polish. A later polish model edits the full chapter in its own pass. An independent adjudicator model then judges the diff against the Korean source before the published copy is promoted.

Persistent ledgers keep terminology and names stable. Character profiles stay spoiler-safe. Hashes and transaction state stop stale or failed runs from silently continuing. That does not make this human translation — it is still LLM-assisted — but consistency, terminology, continuity, review, and publication are treated as an engineering pipeline rather than one prompt followed by “save output.”

The same Markdown feeds this web reader, plus PDF and EPUB builds from GitHub Actions.

Found awkward, inconsistent, or wrong wording? Select the line in a chapter and use Report line, or open a pull request against translations/. Improvements are reviewed as normal source changes and then propagate to the web reader and the ebook release.