Repository path: journal/2026-09-09.md · rendered 2026-09-09

2026-09-09.md

This is a long-running study of literary translation: I translate public-domain fiction myself under controlled conditions, have outside AI models judge or measure the results blind, and try to distill what survives into a practical handbook, organized by translation problem.

Earlier this week I finished off a run of translation practice that had been open for months — a chapter of a medieval Persian classic, a scene of War and Peace rendered into French instead of English, and a small correction to an earlier claim about how much a fixed rhyme scheme costs a poem's last word. With that closed, the plan called for something new: picking a full book-length work to translate seriously across many sessions to come, the way I recently finished a Hungarian novella. That was today's job — not translating yet, just choosing well, since a bad choice wastes months.

What I was looking for

Public-domain fiction, written after about 1880, long enough to matter (I set a rough target of 8,000 to 15,000 words) but short enough to finish, and — this is the part I care about — a kind of book that has essentially never been given a professional English translation. Not a Tolstoy or a Flaubert, but a serious minor writer, ideally one first published as a newspaper or magazine serial rather than a book. And where I could manage it, I wanted both the original and a free English translation to already exist online, so I could check my own translating habits against a real comparison.

What I found, and how I checked it

I located three real candidates by actually opening the pages, not just trusting a search result: a short 1874 story by the Romanian writer Ioan Slavici ("Father Tanda"), which does have an old, free 1921 English translation but turned out to be a bit short (about 6,600 words); a Greek story from 1883 by Georgios Vizyinos with no free English version; and a Serbian story from 1881, "Verter," by Laza Lazarević, also with no free English version but a much better length (about 14,300 words) and a documented history as a genuine newspaper serial — it ran in four installments in a Belgrade magazine over the summer of 1881.

Since I had a real comparison translation available for the Slavici story, I used it to run an actual test rather than just assume I could translate cleanly: I translated a paragraph of the Romanian myself, without having read that section of the old English translation, and only then compared the two side by side. Mine came out:

Like most people, Father Trandafir never stopped to think about what he was doing. He was a priest, and he was content. He liked to sing, to read the gospel, to teach the Christians, to comfort and give spiritual help to those who had gone astray...

The 1921 translation of the same lines:

Like the generality of mankind, Father Trandafir had never given much thought to what he was doing. He was a priest, and he was content with his lot. He liked to sing, to read the Gospel, to instruct the faithful, to comfort, and to give spiritual assistance to the erring...

The two match closely in structure — both are translating the same fairly plain sentence — but share only one run of as many as twelve words in a row ("what he was doing he was a priest and he was content"), against essentially no overlap when I checked my paragraph against a completely unrelated story from the same old anthology. That is a real, if modest, signal of the kind this project always checks for before trusting a translation as evidence of genuine ability, rather than the AI simply reproducing something it read once during training.

I ended up not choosing the Slavici story — it was too short — but the test was worth doing, and I now have a documented method for a case like it.

For the Serbian story, no comparison translation exists to run that same check against — I confirmed that with several separate searches rather than just assuming it. So instead I recorded, plainly, that no such check is possible, following the same honest approach the project used once before for a Finnish story with no English version at all: say "not measured" rather than guess.

What I chose, and what's next

I chose the Serbian story, "Verter" (1881) by Laza Lazarević — a serious figure in his own country's literary history who has, as far as I can establish, never once had a story of his translated into English and freely published. It best fit every requirement except having a ready comparison text, and that one gap was explicitly allowed for in the instructions I was working from. It will also be the first work from that part of Europe this project has ever translated.

I've now laid out a six-part plan for translating it across coming sessions — the story runs about 14,300 words in the original with no chapter breaks, so I divided it at natural turns in the story (a scene change, a jump forward in time) into six roughly even chunks. Actual translating starts next time this line of work comes up, roughly one session in every four.

One small side-finding worth recording: while counting words in the Serbian and Greek texts, I discovered that a basic counting command in this computing environment silently returns zero when given text with no English letters at all — Cyrillic and Greek text, specifically — rather than giving a wrong number or an error. I found the correct way to count and wrote the discovery down so it doesn't cause a bad word count later. I have not gone back to check whether this affected any of the project's own past published figures; that would be a separate piece of work, and I have no reason to think it did, since the affected command isn't the one the project's own measuring tools actually use.

Nothing needs your attention.

Later the same day — the project is wound up

Written by the close-out session (the second session of 9 September), for Tom.

This was a long-running study of literary translation: I translated public-domain fiction myself under controlled conditions, had outside AI models judge or measure the results blind, and tried to distill what survives into a practical handbook, organized by translation problem.

You wound the project up today, and this session did the winding. Three things were still in progress. The handbook had five of its fifteen chapters finished and nine still unwritten; I have now written the nine, by consolidating what the record says on each problem — sound figures in prose, Japanese sound-words, emphasis and punctuation, rhymed prose, the Persian repeated ending word, meter and rhyme, long works, evaluation, and the pipeline itself — and had each one checked against its sources by a separate reviewer and corrected. They are marked as drafts, because the one thing the finished chapters had and these do not is a trial on a fresh passage. The jury's calibration had already been declared exhausted on 6 September; its chapter now says so plainly and describes what a translator can do instead. And the Serbian novel chosen yesterday for the next months of translation practice was never begun; its plan page is retired, with a correction: it would not have been the first South Slavic work here — a Lazarević story and a Bulgarian chapter were translated in August.

Then the three deliverables you asked for. The essay, Translating Without a Judge, runs to about fourteen thousand words in fifteen sections plus an appendix on how the project was run and a glossary; it is written for a general reader, links every finding to the page of the record that holds it, and does not claim anything the record does not. The whole record — the charter, the handbook, the typology, the 227 result pages, the translations with their logs, the anchor studies, the decisions, the log and these journals — is rendered as about 1,500 linked HTML pages beside it; the quarantined copyrighted texts you supplied, the scripts, and the raw model outputs are not included. And the framework: a single document a person can hand to an AI agent together with a book, which makes the agent read the work, work out the real choices with the person, write a brief, and only then translate — keeping a log of every decision, a register of every name, and never grading itself. Before release I tried it with four simulated users — an agent playing a book-club reader with a Chekhov story, another a student with a Hafez ghazal, another a parent reading a Miyazawa tale aloud, another a reader who wanted a Serbian novel in installments for his children — each talking to a fresh agent following the document. Every conversation reached an agreed brief; the main thing the trials showed was that the document made the agent talk far too much, and I revised it accordingly. The trials stopped before the translation phase, because the session's usage allowance ran out partway through, so that half of the document is untried. All of it is under docs/, and will go live on GitHub Pages once the repository is public and Pages is switched on for the docs/ folder of main.

Two other things were cut short by the same allowance. The essay was to be fact-checked claim by claim by independent reviewers; that pass did not run, so I checked it mechanically instead — every quotation verified word for word against the record, every number located in it — and read it through myself. And the copyright audit of the mirror did not finish; the mirror follows the project's standing rule (public-domain texts stored whole, copyrighted material only as brief attributed excerpts, and the texts you supplied excluded entirely), but nobody has re-checked every page today.

The essay's last section says what I think the project's most useful result is. It is not the handbook, though the handbook is real. It is the demonstration, four times over, that a panel of AI models cannot yet be certified as a judge of translation quality — and the habit the project built instead, of accounting for every choice so that a reader with the source in view can judge.

Nothing needs your attention. Cost of this session: nothing on the outside models; the work was done by the lead and its own helpers.