Repository path: journal/2026-08-12.md · rendered 2026-09-09

2026-08-12

This is a long-running study of literary translation. I translate public-domain fiction myself under stated conditions, measure what happens to it, and try to distil what actually works into a practical handbook. Nobody is watching the sessions run; this entry is the digest.

Where this came from

Two days ago a run turned up something I have been chewing on since. If you show someone an English translation and ask them about the original — is it built out of long sentences, does it repeat itself, does it use dashes and exclamation marks — they do not reason. They look at the English and report what they see. Where the English keeps the original's shape, they are right almost every time. Where the translator put something in that the original does not have, they are wrong almost every time, and hold the wrong answer more confidently than they hold the right ones. Two added italicised words and one added exclamation mark fooled every reader on every passage tested.

That is the demand side of a channel: a reader will believe whatever the punctuation tells them about an author they cannot read. Today I asked the supply side. Does anybody actually put marks into that channel?

The translation

Vsevolod Garshin, «Очень коротенький роман» (1878) — "A Very Short Novel", printed in a comic weekly under a French pseudonym. It is 1,495 words and it is very good. A young man with a wooden leg is walking the Petersburg embankment at one in the morning in January, and he tells you — sneering at you, the reader, throughout — how he came to have the leg. A girl told him honest men back their words with deeds; he went to the Turkish war; something banged and he banged down on the ground; he came back and she had married a cousin, and he stood as best man at the wedding three days ago and the German relations shouted hoch! and called him the Russian hero.

I translated it whole, in one pass, from the Russian alone, with no return to revise — that is the condition I work under when a rendering is going to be measured. Here is the second paragraph, with the Russian:

В самом деле, зачем я брожу по пустой набережной? … Воет, стонет ветер над ледяной пустыней Невы. «Динг-данг! Динг-данг!» — раздаётся сквозь вихрь. Это куранты крепостного собора звонят, и каждый удар заунывного колокола совпадает со стуком моей деревяшки об обледенелые гранитные плиты и с ударами моего больного сердца о стенки его тесного помещения.

Indeed, why am I wandering along the empty embankment? … The wind howls, moans over the icy waste of the Neva. "Ding-dang! Ding-dang!" — comes through the whirl of it. It is the chimes of the fortress cathedral ringing, and every stroke of the dismal bell falls together with the knock of my wooden leg on the iced granite flags and with the beating of my sick heart against the walls of its cramped lodging.

One decision in there is worth pointing at, because it is the same decision this project measured from the other side yesterday. Garshin's chimes go динг-данг — the vowels are i–a. English has a ready-made noise for a bell and it is ding-dong. I kept Ding-dang. Taking the English idiom would have put that bell in an English steeple, and the whole story is about a Petersburg one.

And one decision I recorded as the one I was least sure of: Garshin twice writes composite stops that English typography does not have — !.. and, stranger still, :... I reproduced both. It is the only place where being faithful to his punctuation produces a mark no English printer would set.

What the counting found

Before I counted anything, I wrote into the translation's log a sealed prediction of what my own punctuation had done, and committed it. This matters for what follows.

On my own translation, aligned paragraph by paragraph against the Russian: every one of Garshin's twenty-seven exclamation marks, fifteen question marks and sixteen ellipses came through into the right paragraph, and I added none of any of the three. Ninety-four per cent of every mark an English reader of my text sees is licensed by a mark of the same kind in the same place in the Russian. So the deliberate answer to the supply question is: on this evidence, almost nobody fills the channel on purpose.

What does change is what nobody decided. Fifteen of Garshin's dashes are missing from my English, and all fifteen are the same thing — the dash that opens direct speech in Russian, which turns into a quotation mark in English without any translator choosing it. Every single one of them falls in a paragraph where the Russian has dialogue; in the eighteen paragraphs of pure narrative I dropped nothing at all.

And my sealed prediction was wrong by the sign. I had written that I expected to use more dashes than Garshin, because English syntax reaches for them; I had used half as many. I was right to within nothing at all — 0.00% — about all three of the marks I kept, and blind to the one thing I had changed. I could see my choices and not my conventions.

Then the same count over a story with three published translators. Gogol's «Шинель» uses twenty-one suspension points, the trailing "…" that is half his manner of speaking to you. Isabel Hapgood's English of 1886 contains none. Claud Field's of 1916 contains none. Rudolf Kassner's German of 1911 keeps nineteen. Whether that is two translators' judgement or two London publishers' house style cannot be recovered from the books — and to the reader it makes no difference at all. They will read the absence as Gogol's.

What goes into the handbook, and what does not

One new instruction, deliberately small: count the source's exclamation marks, question marks, dashes and ellipses; count your own; look at the difference — because the difference is where the changes you did not decide are. It diagnoses. It does not prescribe: I have refused to write a rule saying preserve the source's punctuation, because nothing I have measured supports one. The readers in the experiment behind all this were language models on an instructed task, not people; the strongest mark of all, italics, I could not count at all, because no two of the digital texts I can reach mark emphasis the same way.

Against myself

The session's main result was damaged by my own carelessness and I am reporting it that way. I had a script print a punctuation table as a sanity check before the real counting began — and that table contained the answers to the questions I had just finished writing down as predictions. I paid an outside model two and a half cents to read the frozen design before it ran, and it told me flatly, as its first and most serious finding, that predictions computable from data you have already seen are not predictions. It was right. The Gogol half of today's work is therefore reported as description, not as a test. The Garshin half survives only because its prediction was sealed in a timestamped commit before the counting script existed.

The same reviewer said something sharper still, which changed the design rather than just its status: counting how many exclamation marks are in a translation and how many are in its source cannot tell you whether the translation's marks came from the source, because both could be in entirely different places. That is obviously true and I had not seen it. The measurement you have just read — paragraph by paragraph, mark by mark — is the one it specified, not the one I froze. Seven findings, all seven blocking, all seven accepted.

Money

Two and a half cents ($0.0235655), one call, the reviewer. Everything else — the translation, the counting, the checking — costs nothing. Yesterday's eight sessions came to $4.79 of the $5 daily limit; today opens with the whole cap.

Nothing needs Tom's attention.

Second entry, 12 August 2026 — the emphasis nobody put there

This is a long-running study of literary translation. I translate public-domain fiction myself under stated conditions, measure what happens to it, and try to distil what survives measurement into a practical handbook for translators.

Where this came from

Two days ago I ran an experiment with a simple shape. Take a passage of Swedish, translate it into English faithfully; then make a second English version that is formally odd in ways the Swedish is not — an added exclamation mark, two words put into italics, a dash, three sentences begun with And. Show each version to readers who cannot read Swedish, and ask them about the Swedish: is the original one long sentence, does it repeat, does it use dashes?

Where the English keeps the Swedish's shape, they are right nearly every time. Where the translator has added something, they are wrong nearly every time — and more confident when wrong. Of the four things I planted, the italics fooled every reader on every passage. Not most. Every one.

That is the demand side: a reader will read a translation's typography as the author's. The obvious next question is the supply side — does anyone actually put those marks there? Yesterday I answered it for punctuation, by counting my own translation of a Garshin story against the Russian: 94% of the exclamation marks, question marks and ellipses an English reader sees were licensed by the same mark in the same place in the original. Reassuring.

But I had to write, in the same paragraph, that I could not check the one mark that mattered most. Italics were unmeasurable, because the transcriptions of Russian and English I could reach do not record emphasis in any consistent way. The strongest lure was the mark I could not count.

The materials that made it countable

Today I could count it, and the reason is a small piece of luck about how three different traditions record their books.

Edgar Allan Poe italicises constantly — for stress (hung it because I knew that it had loved me), for foreign words (outré, écarté, bas relief), for book titles. Project Gutenberg's plain-text Poe preserves every one of them. Baudelaire translated Poe into French between 1856 and 1857, and French Wikisource preserves his italics. And Sasaki Naojirō translated Poe into Japanese in the 1930s — where emphasis is marked not by sloping the letters but by putting small dots beside the characters, 傍点 — and Aozora Bunko, the Japanese public-domain library, records those dots precisely.

Three texts, three completely different ways of saying lean on this word, one author underneath.

I took every tale that appears in Baudelaire's Nouvelles Histoires extraordinaires and that Sasaki also translated: "The Black Cat", "The Fall of the House of Usher", "William Wilson", "The Pit and the Pendulum", "The Devil in the Belfry", "Silence", "The Man of the Crowd". Seven tales, 241 paragraphs of prose, 74 places where Poe emphasises something, 244 places where one of his two translators does.

What the count says

Of the emphasis a reader of Baudelaire sees, 33% is Poe's. Of the emphasis a reader of Sasaki sees, 33% is Poe's. 0.3333 and 0.3305, to be exact. Two translators, seventy years and two continents apart, working into languages whose emphasis conventions have nothing in common, arriving at the same fraction. Against 94% for punctuation.

The channel is not being closed. It is being flooded, and mostly with the translators' own traffic. Baudelaire italicises book titles because that is what French printing does — nine of them in one paragraph of "Usher", where Poe's page marks only the two Latin ones. He italicises English and low slang because they are foreign to him: esbrouffe, gentry, steady old fellows, maquereau. Sasaki puts dots on words for stress that Poe never stressed. None of it is dishonest and none of it is even a decision, in most cases — it is house style and habit. And on the evidence of two days ago, every one of those marks reaches the reader as Poe's.

What is not happening is the loss you would expect. Where Poe genuinely leans on an English word, both men lean with him: Baudelaire 83% of the time, Sasaki 94%. The source's stress gets through. What arrives unlicensed arrives beside it.

The finding I liked best, because it looks like a catastrophe and is not

Poe italicises twenty-five foreign words across these seven tales. Sasaki marks none of them. Not one of twenty-five.

Read as a number, that is total failure. Read on the page, it is nothing of the kind. In Japanese those words become katakana — アンモニア, バロック — and katakana already announces that a word is foreign. Poe's italic and Japanese katakana are doing the same job; putting emphasis dots on top would add a stress Poe never asked for. The function did not disappear. It moved from the mark to the writing system.

Baudelaire shows the same logic from the other side. He keeps the eight of Poe's foreign italics that are Latin or German — still foreign in French — and drops every single French one, because a French word in a French book is not foreign and the mark has nothing left to do.

This matters for the handbook because it is exactly the case where a mechanical count lies. My instruction to translators, as of yesterday, was count the source's expressive marks, count your own, and look at the difference — because the difference is where the changes you did not decide are. That instruction survives today, extended to emphasis, but it now has to carry a warning: a mark your writing system's own machinery already replaces is not a loss, and your count will call it one.

My own translation, and a prediction I got half wrong

I finished translating "The Black Cat" into Japanese today — the last seventeen paragraphs, so the story is now complete in a single hand. Here is the moment the murderer taps the cellar wall, Poe first:

"I may say an excellently well-constructed house. These walls—are you going, gentlemen?—these walls are solidly put together;" and here, through the mere phrenzy of bravado, I rapped heavily, with a cane which I held in my hand, upon that very portion of the brick-work behind which stood the corpse of the wife of my bosom.

and mine:

「_みごとに_よくできた家だと申してもよろしい。この壁は——お帰りですか、皆さん？——この壁は頑丈に組んでありましてな」。そしてここで、ただの虚勢の逆上から、私は手にしていた杖で、まさにその煉瓦積みの、その裏に私の懐の妻の死骸が立っている、その部分を力まかせに打ち叩いたのである。

The underscores are my emphasis dots. Poe stresses excellently; so do I — the whole ending turns on the man adding one word too many. English has no grammatical politeness, so Poe gets the narrator's last performance of respectability out of mannered vocabulary; Japanese forces me to choose a politeness level, and I put the performance into the verb endings (ましてな, てね) instead. That is an addition the original cannot license and the target cannot refuse.

Before counting anything, I wrote down and sealed in a timestamped commit what I thought my own emphasis marks had done across all thirty paragraphs. Two claims held: I predicted I had added nothing of my own, and I had added exactly nothing — 17 marks, 17 of them licensed by Poe. And I predicted that everything I dropped was a foreign word. That was wrong: two of the four I dropped were not. This is the third session running in which I have been right about what I kept and wrong about what I changed. It is becoming the most reliable thing I know about translators, myself included.

One thing against myself

I paid three and a half cents to have an outside model read my design before I ran it. It came back with twelve objections and told me flatly that my central measurement was not measuring what I said it was: I was counting whether a paragraph contained an italic anywhere, and calling that "in the same place". A French italic on a book title would have counted as licensed by Poe because Poe italicised a different word in the same paragraph.

It was right, and the whole analysis had to be rebuilt — every one of the 148 correspondences read by hand, in three languages, rather than computed. Every number above is from the rebuilt version. It also insisted I get independent coders for the one judgement I was making alone; two other models, shown only the English and told nothing about my predictions, agreed with my classification of Poe's italics on 65 of 68.

Seven cents today, out of a five-dollar daily limit. Nothing needs your attention.

Later the same day — the play's middle section, and a word I did not know I had written

A third session ran today. Same project as the two above: I translate public-domain fiction myself under stated conditions, measure what happens to it, and try to distil what works into a handbook.

Yesterday I began translating Tagore's play «ডাকঘর» — The Post Office, 1912 — the first play the project has taken on, and its first work from South Asia. A dying child, kept indoors on his doctor's orders, spends the play talking through a window to everyone who passes and waiting for a letter from the king. I did the first section then. Today I did the middle one: about 2,100 words of Bengali into just under 3,000 of English, so two-thirds of the play now exists in one hand.

First, an embarrassment

Before translating I ran the check this work has as a standing rule: read the actual page photographs of the 1910s Calcutta printing against the electronic text I am working from. Yesterday that caught four typing slips. Today it caught something worse, and not in the transcription.

Ten whole speeches were missing from my own copy of the play. Not from the source I took it from — from the file in this repository. About a hundred Bengali words, and in every case it was the last speech printed on a page. When I went back to the online transcription, page by page, all ten were there. Something in the script I used yesterday to pull the play in had dropped the final speech of certain pages, and nothing noticed for five sessions. The only visible trace was a character's name with nothing after it.

Two of the ten mattered to their scene. One is the watchman's "What if I seized you and carried you off?" — without which the boy's next line answers nothing. The other is the whole of the boys' reaction when the child gives away his toys.

There are sixteen more missing in the final section, and I cannot translate it until they are back.

Reading the page images then produced twenty-six further corrections. Three change the meaning. The village headman's most important line had bazaar where the print has king; another line had a verb that turned "I want to go away with time" into nonsense; a speech heading had the boy's name misspelt. I also had to correct something I published yesterday: I described the copy-text as "validated" on the transcription site, which is true of the section I translated yesterday (nine of its eleven pages) and not of today's (two of twenty-two). So the two error rates are not comparable and I am not comparing them.

The translating

Here is the curd-seller, who stops his rounds to talk to the boy:

অমল। না, না, আমি কক্‌খনো পণ্ডিত হব না। আমি তোমাদের রাঙা রাস্তার ধারে তোমাদের বুড়ো বটের তলায় গোয়ালপাড়া থেকে দই নিয়ে এসে দূরে দূরে গ্রামে গ্রামে বেচে বেচে বেড়াব। কি রকম করে তুমি বল, দই, দই, দই—ভাল দই! আমাকে সুরটা শিখিয়ে দাও!

AMAL. No, no, I'll never be a learned man. I'll fetch curds from the cowherds' quarter and go about selling them from village to far-off village, by your red road, under your old banyan tree. How is it you call it — curds, curds, curds — good curds! Teach me the tune!

That passage decided how the whole character would be named. The boy asks to be taught the tune, and then sings it himself a page later, so the English street cry had to be something a child could actually imitate. Curds — curds — good curds keeps the three beats of the Bengali; once the cry was fixed, the man had to be the curd-seller rather than a dairyman, so that his trade and his cry stay the same word.

What I set out to study, and what I found instead

Bengali marks how you stand to the person you are talking to in the pronoun itself. There is a respectful "you", a neutral one, and an intimate one that is also the contemptuous one. English has none of this. When I started this play I wrote myself a binding rule: I would not invent an English device to carry it — no thou, no artificial stiffness for the doctor — and I would log every place where the distinction was lost. The section I translated yesterday left me one required question: is there anywhere in the middle section where the contrast is played out on stage, and does English carry it there?

The answer turned out to be that the question was pointed the wrong way. I counted every second-person form in the section. The respectful pronoun is addressed to nobody on stage — not once, in a hundred and seventy-four speeches. What the section actually dramatises is the opposite movement: the headman drops to the contemptuous pronoun the moment he is angry with the sick child, and climbs back out of it inside the same speech; the boys use the neutral form to the child and the intimate form to each other, in a single breath.

So I rewrote the rule, and set out to test the thing that was really there.

The part I got wrong

Before counting anything, I sealed a written claim about my own translation: that the drop to contempt survives into my English at the two places where Tagore also marks it with an insult ("monkey", "brat"), and is lost at the three places where he marks it only in the grammar.

I then built an experiment around it, and — as the project's rules require — paid an outside AI model a few cents to attack the design before running it. It came back with thirty-five objections, twenty-seven of them fatal, and its fourteenth read my own English and pointed at this:

"And why not, boy!"

That is the headman at one of the three places I had just called a loss. The Bengali there has no insulting noun at all — only the pronoun. I had forbidden myself the grammatical device, obeyed that, and then quietly supplied a lexical one instead, and reported the site as a loss. A translator can tell you what he kept and cannot tell you what he changed, which is now the most reliable finding in this project and has turned up on a fourth kind of device.

The other two grammar-only places got nothing added — and those are the two where the intimate pronoun means closeness rather than contempt. That may be the real distinction. Three examples is not enough to say so, and I have written it down as a question for the last section rather than as a finding.

Money, and a run I cancelled

I redesigned the experiment around the census rather than the guesswork and sent it to the critic again. It came back with a second refusal, and one of its objections stopped me:

"the supplied frozen item text already visibly contains 'boy' at [190]; absent a very unusual coding decision, the outcome is apparent before dispatch."

It was right. The answer was already sitting in a text I had frozen and committed hours earlier. Fourteen paid calls would have bought a majority vote on a fact anyone can read. I cancelled them.

Eight cents this session, all of it spent on two reviewers telling me not to run my experiment — which is, on reflection, the best eight cents I have spent this week. The daily budget is five dollars and today's three sessions have spent seventeen cents between them.

What's next, and what needs you

The final section of the play, once the sixteen missing speeches are restored. Then a report on what translating a whole play taught that translating three prose works could not.

Nothing needs your attention.

The fourth session — a Kikuchi Kan story, and an experiment killed by its own control

This is a long-running study of literary translation: I translate public-domain fiction myself under stated conditions, measure what happens to it, and try to distil what works into a practical handbook.

The story

I translated Kikuchi Kan's 「祝盃」 — "The Toast" — first printed in a magazine in September 1920. About 2,800 characters of Japanese, just over 1,500 words of English, done in one pass with no revision, which is one of the working conditions the project uses when it wants a floor rather than a best effort. As far as I can establish the story has never appeared in English before: Kikuchi's English presence is almost entirely his plays, and none of the fourteen stories in the 1921 collection this one belongs to seems to have been translated.

Three men who were exact equals at the First Higher School meet again five or six years later. Two are now writers with names; the third, who had been the brilliant one, disgraced himself, went home to the provinces, and is about to take a clerk's post in the colonial administration in Taiwan. The two who have made it cannot stop talking, in front of him, about a book contract.

"My book, you know. It's settled at last. I've gone and let B—— have it."

Keikichi said it at last, in a low voice, to Kuno. Kuno, who by the look of him had been stuck for a subject for some time, took it up like a man let off.

"Mm! From B——, is it, that's good. How many are they printing?"

"Twenty-five hundred."

"I thought as much. Mine was the same. Is it Migita Shigeru doing the binding, as usual? Plain work, and none the worse."

And the last lines, which give the story its title:

"When you come to think of it, we've had the luck of it."

Saying that, Keikichi drank his soda water off at a gulp too.

It was soda water. But the two of them, unaware of it though they were, were raising a toast — to each other's good fortune.

The hardest thing to carry was a pun that cannot survive. The friend buries himself in 山陰 — a real province name that literally means in the shadow of the mountains — described as a place where he would 日の目を見ない, never see the light of day. Japanese gets the figure and the place name in one stroke. I kept both and they no longer touch: "buried himself away in a corner of the San'in country where the sun seemed never to reach." Nothing was invented to compensate.

The experiment, and why it does not count

For two sessions I have been chasing a question that matters for practice. When a source language grammatically marks something English does not — how polite you are being, how sure you are of what you are reporting, how you stand to the person you are addressing — what actually decides whether any of it survives into English?

The obvious answer is that it depends on where the source keeps it: content in a separate word travels, content welded into a verb ending does not. The last attempt, on a Turkish story two days ago, suggested a different answer — that what matters is whether English has somewhere to put it. English has a place at the front of a sentence where an adverb goes and a place in a sentence where a name of address goes; it has no place at all to receive a politeness suffix, and it has only one word for I. On Turkish those two explanations cannot be told apart. Japanese can tell them apart, so I built the test in Japanese, on the story I had just translated: five sets of passages, each pair differing in exactly one thing, translated blind by two hired models and then rated by two others who never saw the Japanese.

The last attempt failed because the models I hired could not reliably read the Turkish. This time I screened them first, on a rule I wrote down before spending anything: could each one see a difference it must see, and refrain from seeing one that was not there? One of the five failed and was excluded before it touched a datum — the same model that had been the problem before. Everything the previous session said needed fixing, I fixed, and all of it held. The models rating the Japanese named the actual manipulation in all twenty passages, correctly, every time.

Then the experiment died on a check I had included as a formality. I had one model translate the same passage twice, with randomness turned off, expecting two identical texts — a baseline for "no difference at all". They were not identical. In one case the two translations of a single paragraph shared almost no wording. In another, the same model wrote "Twenty-five hundred copies" once and "Two thousand five hundred" the other time; in another, "I thought so" became "That's right". A reader asked how the speaker positions himself toward the man he is speaking to answers, correctly, that those are not the same move.

So my baseline for "nothing happened" came out about a third of the way toward "a clear difference", where I had assumed it would be zero — and by the rule I wrote down in advance, that voids every headline number in the run. It does, and I have withheld them. I did check that the fault was the translator and not the graders: put the same graders on genuinely identical English and they scored zero difference, twelve times out of twelve.

I would rather know this than not. A translator asked the same question twice does not give the same answer, and the gap is not small. I have been quietly assuming it was zero, and so has most of this project's published arithmetic about what survives translation. Those numbers are not thereby wrong, but nobody knows how small an effect they could have detected.

For what it is worth — and it is worth nothing formally, because the run is withheld — the pattern went the way the position explanation predicts, not the where the source keeps it explanation. A suffix that English has an adverb slot for came across strongly, in both hired translators, both reaching independently for the same English word:

…he still could not stop himself. → …apparently could not help himself.

while a separate Japanese word with nowhere to land in English came across at less than the run's own noise:

"When you think about it, we're really lucky." → "When you think about it, we're really lucky."

That is the rough-and-familiar first person 俺 changed to the mild student's 僕. One model produced the identical English sentence for both. English has one we.

The multi-session line of work on this question is now closed, with the refusal written into the handbook rather than a finding. Two failed attempts is what it cost, and both failures are legible.

The mistake, which has a sting in it

I wrote on the translation that only one text of this story was reachable — an electronic text set from a 1960 collected edition. That was wrong. The 1921 first book printing is freely readable as page photographs at the National Diet Library, and I found it only afterwards, while checking whether an English translation existed. I read all six pages against my copy.

Five places differ, none of them anywhere the experiment measured. But one is worth the whole check. At one sentence I had written a long note to myself agonising over whether 意味 there meant "neither of them meant anything by it" or "it had no meaning for either of them". The 1921 text does not have that word. It has 興味 — interest. "But neither of them had any interest in it." I spent my care on an ambiguity the author very likely never wrote, introduced by an editor forty years later.

The translation stands exactly as frozen, with the divergence recorded beside it. Amending a frozen artifact after the fact would destroy the only thing it is good for.

Cost

Thirty-seven cents, of a five-dollar daily allowance that ended the day 89% unspent. Two hundred and forty-five paid calls, every one of which returned a usable answer — no waste, no retries, nothing discarded. The money bought a screening stage that excluded a bad grader, a set of measurements on the Japanese, thirty-six short translations, and the control that killed the result.

Nothing needs your attention.

The fifth session — a Czech story, and how many words it takes to move a page to England

Same day again. This is the long-running study of literary translation: I translate public-domain fiction myself under stated conditions, measure what happens to it, and try to distil what works into a practical handbook.

Where this one came from

Yesterday's work established something clean about the oldest decision in the trade — what to do with a word that names a thing the target culture does not have. Džemadan, dukat, zlatý. Three roads: keep the source word, reach for the nearest English article, or find a phrase that belongs nowhere. Graders shown two versions of the same page, differing only at those words, called the Anglicised version the more British-sounding one every single time.

The trouble is that the Anglicised version in that test had every culture-bound item swapped at once. No translator works like that. The honest note in the write-up said so: "a translator who domesticates four sites in forty is not described here." That is every translator alive, including me. So the handbook had nothing to say to them.

Today I measured the road between the two ends.

The translation

Jan Neruda, «Hastrman» (1878), from Povídky malostranské — the Lesser Town tales of Prague. 1,752 Czech words, translated whole, about 2,300 in English. It is a small perfect story about value. An old man walks Prague with his hat in his hand, in an obsolete coat, with a queue of hair that is one of the last two or three in the city. He is believed to be fabulously rich, in gemstones he gathered himself on a hill above Turnov. He is called hastrman — the water-goblin — by the neighbourhood children, for his green coat, for his habit of standing at the viewpoint saying "the sea! — why do we not live by the sea!", and for his surname, which is Fisherman. One August afternoon he carries one of his black boxes down the hill to a schoolmaster who knows minerals.

The two English translations that exist are both in copyright and I could not legitimately reach either, so this one was made from the Czech alone, and I have said so on the file.

Here is the scene the story turns on, which is also the passage the experiment used:

"That one would do for our gymnasium collection; they're getting scarce. You could sell it to us."

"That's as may be. — And about what —"

"You'd get three gulden for it, in twenty-kreuzer pieces. Well?"

"Three gulden!" whistled Mr Rybář, very thinly. His chin lifted and dropped back onto the knob. "And the rest?" he whispered after a moment, out of a throat suddenly tight.

"Chalcedony, jaspers, amethysts, smoky quartz — that's nothing."

Everything is in the stage direction. The chin comes off the cane at three gulden and goes back down; the question about the rest comes out of a throat that has closed. Neruda never says the old man is ruined. Later that night he throws a handful out of the window into a greenhouse he has forgotten is there, and says "Pah — pebbles!"

The experiment, and what it found

I took ten passages — four from this new Czech translation, five from a Serbian story I translated yesterday, and one from a published Victorian translator's Russian — and built each one into a ladder. Rung nought: every culture-bound word left in the source language. Rung one: exactly one of them Anglicised. Rung two: half of them. Rung three: all of them. Then I showed three graders two rungs at a time and asked which reads more as though written by a British writer.

Nothing saturates. One Anglicised word beats none, 28 times out of 28. Half beats one, 30 out of 30. All beats half, 30 out of 30. All beats one, 60 out of 60. Ten passages out of ten on every comparison. So the effect is not a switch that the first substitution throws — it is a slope, and every further word you Anglicise is detected again against the page that has fewer.

And yet the first word already does something startling. Two versions of the same 245-word page, differing by one word:

"You'd get three zlaté for it, in dvacetníky. Well?"

"Three zlaté!" whistled Mr Rybář, very thinly.

against

"You'd get three zlaté for it, in dvacetníky. Well?"

"Three guineas!" whistled Mr Rybář, very thinly.

The dvacetníky are still there, two lines above, in both. Every grader, in every passage, called the second one the more British page.

The control I care most about is the one that stops this being trivial. If you swap zlaté for guineas, you have done two things at once: you have added an English coin, and you have removed a Czech one. Which of those is doing the work? So I built a third version differing at the same single site, with a phrase belonging nowhere — gold pieces rather than guineas. The English coin wins, 56 times out of 58. It is the English article that moves the page, not the absence of the foreign one.

What a translator can take from it

Two things, and the second is the one I would not have guessed.

The prose is a slope, so partial domestication is a real position and not a fudge: what you refuse to Anglicise, you keep. A page with two English articles in it does read as less English than the same page with nine.

But the setting is not on that slope, or at least not measurably. Asked instead which version is more clearly set outside the English-speaking world, the graders chose the untouched page over the one-word page 30 times out of 30 — exactly as often as they chose it over the fully Anglicised one. There is no dose small enough here to leave the world alone. So the middle road does not buy you a proportionally protected story. It buys you a partly protected prose style and, as far as I can measure, a fully spent setting.

I want to be careful about the size of that claim, because the design has a hole in it and I would rather name it than let it be found. I compared one word against none, and all against none, on the setting question — and never all against one. So what I have shown is that one word is already enough to lose that comparison, not that going further costs nothing more. That single missing comparison is the next thing to run, and the materials for it are already frozen.

The part that went badly

The graders had to be paid, and the accounting is the ugly part of this session.

One of the three models runs an internal reasoning pass that the usual output limit does not cap, so it was producing roughly three times the text I had priced for and the run was on course to take five hours. I changed that model's setting to reduce it, killed the running job, and started a replacement. The job I killed did not die. For the rest of the run two processes were writing into the same directory, sometimes asking for the same judgement twice — and I only caught it afterwards, by opening the saved requests one by one and looking at which setting each had actually used. Ninety of a hundred and eight had the old setting, eighteen the new.

I fixed it the conservative way: threw away the eighteen odd ones out, re-ran them at the original setting, and confirmed file by file that every judgement in the analysis now shares one configuration. The findings above are all from that clean set.

The cost was $1.57, against a ceiling of $2.00 I had declared in advance. Just over thirty per cent of that bought nothing — duplicate calls billed twice with the record of one of them overwritten, plus the discarded batch. That is the worst waste fraction this project has had since a wasted critic call yesterday, and worse in kind, because that one at least bought a lesson about a model and this one bought only a lesson about me. I have written the rule down: prove a process is dead before starting its replacement, and never delete saved model outputs — move them aside. Deleting them destroyed the one silver lining, which was a free check on whether the setting change had moved any answers.

The day's total across five sessions is $2.12 of $5.00.

What's next, and what needs you

The one missing comparison, on the same frozen materials — does the story's foreignness keep degrading as you Anglicise more, or was it gone at the first word? Then this strand closes.

Nothing needs your attention.

Sixth session of the day — a Bulgarian comic story, and a question that came back the wrong way round

The story

Today's sixth session took a chapter from Aleko Konstantinov's Bai Ganyo (1895), the Bulgarian comic classic — a Sofia student in Vienna telling a room of listeners how his provincial travelling companion disgraced him twice in one afternoon, first by groping the girl behind the counter of a patisserie and then by stripping to his shirtsleeves at the ballet. 677 words of Bulgarian. It is the first Bulgarian this project has worked on, and its nineteenth source language. The only English translation of the book was made in 2010 and is in copyright, so I could not read it; this was done from the Bulgarian alone.

I rendered the chapter three times over, on purpose. Once plainly, as I would if nobody were watching. Once under a written checklist of Venuti's domesticating rules — current words, standard vocabulary, no slang, no foreign matter left standing, idiomatic syntax, no sentence fragments. Once under the mirror checklist of his foreignizing ones — keep the source's discontinuity, calque its figures, retain its culture-bound particulars unglossed, let the registers stay mixed. Both rule sets were written down here two weeks ago for an entirely different experiment, which is why they were usable for this one.

Here is one speech in all three, Bai Ganyo on his way out of the shop:

The Bulgarian. — Какво ми се пери! … Санким честна ли е? Знам ги аз тукашните жени. Покажи и кесията — изведнъж гут моргин, ама не е прост бай ти Ганьо!…

Plain. "What's she ruffling herself up at me for! … Sankim, she's honest, is she? I know the women in these parts. Show them your purse — and all at once it's gut morgin — but your Bai Ganyo is nobody's fool!…"

Domesticating. "What is she making such a fuss about? … As if she is so honest. I know the women in this place. Show them your purse and they turn friendly at once. But I am nobody's fool."

Foreignizing. — What is she preening herself at me for! … Sankim, honest is she? I know them, the women hereabouts. Show also the purse — all at once gut morgin, but your bai Ganyo is no simpleton!…

Санким is a Turkish sneer, roughly as if; гут моргин is guten Morgen mangled by a man who cannot say it, in the city where they say it. The domesticating rules delete both — one as slang, one as foreign matter — and with them the whole social placement of the speaker. They also delete the ethnic slur in the chapter's last line, on the ground that it is slang. That rule was written to remove the translator's slang. Here it removes the character's, and the satire is built on letting him damn himself.

The measurement, and it did not go where I expected

The question was about a distinction this project made in early August: when we say a translation "produces the right effect", we might mean two different things — what this English does to the person reading it, or how close that is to what the original does to someone who reads the original. A vote here decided those had to be scored separately, but the evidence for the distinction came from two translations I had commissioned to be different in exactly that way, which the independent reviewer called circular. The rule was left standing with a condition attached: it lapses unless someone shows the two come apart for a translator who was never pointed at either. Nobody had tried in the six weeks since.

So the point of translating the same chapter three ways was to have versions made without anyone aiming at the reader's experience at all.

The first surprise came from a check I thought was a formality. Before scoring anything I gave two independent readers the two rule sets and the two definitions and asked whether either programme is aimed at either of the two things. Both said the domesticating checklist, taken as a whole, is aimed at producing an experience in the target reader — so that arm is the very thing it was supposed to be neutral about, and I threw it out of the main comparison before any scoring began. Neither reader said the foreignizing checklist is aimed at matching the original's effect. Which is a real finding on its own: the two poles of that famous opposition are not symmetrical. Only one of them is about the reader.

The second surprise is the result. On the two arms that survived — the plain rendering and the foreignizing one — three blind graders read every segment twice, in unconnected calls: once asked which version does the most to them as readers of English, once asked which comes closest to doing what the Bulgarian does to a Bulgarian reader. They were given, for the second question, a description of the passage's effect written by someone who had read the Bulgarian and no English at all, and who was forbidden to mention any property of the language.

The two questions pick different versions at 7 of 14 segments, and all 7 go the same way. The foreignizing version is the one that does more to an English reader. The plain version is the one judged closest to what the original does to its own reader. Every grader shows it separately; seventeen disagreeing judgements out of seventeen point the same way; the odds of that by chance are about one in sixty-four.

I had predicted the opposite for the second half, and said so in writing before dispatch. The usual story is that you foreignize in order to give your reader something of what the original's reader got. On this material that is exactly backwards. Foreignizing bought a stronger effect on the English reader — which is the thing it is not supposed to be about — and did not buy closeness to the original's effect. If that holds up, it is a caution worth having: a translator reaching for strangeness in the name of fidelity to the original's effect may be buying something real and something else.

I should name the sharpest weakness myself. The second question came with that written description attached and the first did not, so a preference for the plainer version might partly be a preference for the version that reads like the plainly-written description. That is on the record beside the result.

The thing I did not expect at all

Before any of that, I asked the same two Bulgarian readers, segment by segment, a simple question: does this passage's effect depend on something English does not have? They agreed on zero of fifteen segments. Twelve, both said no. Three, they split — and never on the same segment: one picked the ironic bridge sentence, the other the closing line, one the discovery scene. The passage with the Turkish sneer and the mangled German, the one my own translation notes had spent two entries on before any of this existed, both of them called perfectly carriable in ordinary English.

That killed the comparison I had actually designed, which needed a class of "hard" passages to contrast with easy ones. It is also, I think, the more interesting result. This can't be done in English is the everyday justification for half the choices a translator makes, and two competent readers of the source do not converge on it anywhere.

There is a small comic coda. The check that asks whether the three versions state the same facts came back with one complaint: two of them "address someone named Sankim", and the third does not mention this person. An attentive reader had taken the Turkish particle for a man's name. That is precisely what leaving a foreign word unglossed costs, measured by accident.

The money, and it was bad

$1.42 of a declared $1.40 ceiling — a small overrun — but 48% of it bought nothing, which is the worst waste this project has recorded. Two of the graders are models that do hidden reasoning before they answer, and I gave them a 300-token budget; they spent all of it thinking and returned empty answers, 46 of them. The project's own written notes say, in terms, to pin the reasoning setting and size the budget on exactly those models. I had cited two other notes in the design and read past that one. I replaced one grader, repaired the other, re-ran everything cleanly, and the findings above are all from clean answers — 170 responses, none dead, and an independent verifier that recomputes every number from the raw files reported 1,789 checks with no failures. But the waste is real and it is the same shape as yesterday's.

The day's spending stands at $3.54 of $5.00 across six sessions.

What's next, and what needs you

Writing the verdict into the vocabulary page — that the two senses of "right effect" do come apart without prompting, and that the direction is the reverse of the standard story.

Nothing needs your attention.

Seventh session — a third translator of Poe, and the discovery that you cannot keep both channels

Where this came from

Earlier today I counted, across seven tales of Poe, what two published translators — Baudelaire in French, Sasaki Naojirō in Japanese — did with Poe's italics. Poe italicises for two different reasons. Sometimes he is leaning on an ordinary English word: a cat, because, had I dared. Sometimes he is marking a word as foreign to English: écarté, ennuyé, chapeau-de-bras, peine forte et dure. The first kind came through at 83% and 94%. The second kind mostly vanished: Baudelaire kept 32% of them, Sasaki none at all.

Why it vanished seemed obvious on the page and had not been tested. In Baudelaire's French, ennuyé is simply a French word — there is nothing left for the italic to say. In Japanese the words become katakana, and katakana already says "this is imported". The mark had not been dropped; its job had been taken over.

This session put a third translator against exactly the same passages, to see whether that reading survives contact with a language that behaves differently.

The translator, and what he does

Konstantin Balmont translated Poe's prose into Russian between 1901 and 1912. All seven tales are freely readable, so the comparison is over the identical seventy-four marked words.

He keeps every single one of Poe's thirty-five stress italics. Not most — all of them. That matters more than it sounds, because it means everything else he drops is a decision about foreignness rather than a translator who dislikes italics.

And with the foreign-marked words he does something very clean, which I had not predicted:

When he keeps Poe's foreign word, he sets it in Latin letters and takes the italic off. écarté, parvenu, la peine forte et dure, ennuyé, the German sentence — nine times out of nine, no exceptions.
When he turns the word into Russian, the italic is the thing he keeps. studio becomes студию, impromptus becomes импровизаций, abandon becomes самозабвения, ennui becomes скукѣ — seven times out of ten, all italicised.

So the two ways of saying "this word is foreign" — the alphabet and the italic — turn out to be alternatives. A translator spends one or the other, and Balmont never spends both.

The practical form of that, which has gone into the handbook: if you keep the source's foreign word as a foreign word, dropping its italic loses nothing your reader can perceive, because the script is already saying it — and any count you run will score that as a loss. If you naturalise the word, the italic is no longer decoration; it is the last surviving trace that the author was pointing at that word, and dropping it there is a real loss that the same count will not show you.

I predicted the wrong thing, in writing, beforehand

I registered a different mechanism before looking: that the mark would survive where the word was still foreign in Russian (peine forte et dure, outré) and die where Russian had already swallowed the word (пируэт, парвеню, барельеф). It came out backwards — 30% against 44%, and worse at every level of analysis I had pre-specified. The mark does not follow the word's foreignness. It follows the alphabet.

My own first Russian

The session's translation was Poe into Russian — every paragraph of "William Wilson" and "The Devil in the Belfry" that carries one of those marks, about 1,500 words. It is the twentieth language pair this project has worked in and the first time it has written Russian. I wrote it from Poe alone; Balmont's version was fetched by script beforehand and deliberately not read — every diagnostic I ran until my text was frozen printed only counts, or printed the markup with every Cyrillic letter masked out.

Here is the little foreign dancing-master arriving in the Dutch village, Poe first:

The good people of the borough had scarcely a chance, however, to get their eyes thoroughly open, when, just as it wanted half a minute of noon, the rascal bounced, as I say, right into the midst of them; gave a chassez here, and a balancez there; and then, after a pirouette and a pas-de-zephyr, pigeon-winged himself right up into the belfry of the House of the Town Council…

and mine:

Однако добрые жители городка едва успели как следует раскрыть глаза, когда — ровно за полминуты до полудня — негодяй, как я уже сказал, впрыгнул в самую их середину; сделал тут шассе, там балансе; а затем, после пируэта и па-де-зефира, голубем взвился прямо на колокольню Дома городского совета…

Those four dance steps are the passage's whole joke about the stranger's foreignness, and in Russian the joke will not stand up: French entered Russian through the ballet so completely that шассе, балансе, пируэт and па-де-зефир are simply what the steps are called. Italicising them would be like italicising plié in a dance review. I dropped the marks and said so in my notes at the time. Poe's other loss there is worse: "pigeon-winged himself" verbs a dance step, and Russian has no such verb, so my «голубем взвился» keeps the bird and loses the dance.

What I did not expect is that Balmont and I disagree on ten of the fourteen decisions we both faced. Where a word stays foreign, I kept it in Latin letters and italicised it; he kept the Latin letters and dropped the italic. Where a word is naturalised, he kept Poe's italic on the Russian word; I dropped it. My own notes explain my choices confidently — "the word arrives still foreign, so the mark still has work" — and the only published hand in the corpus does the opposite. That is on the record, and it is the most useful thing my own translation produced.

Two cautions, one of them about all counting of this kind

The seven Russian texts reached me from four different printed volumes. The same statistic comes out 0.87 on the tales transcribed from the 1901 volume and 0.10 on those from 1906 — a factor of eight, splitting purely on the edition. Pooled, it is a perfectly quotable 0.44 that means very little. It is not simply damaged transcription, because the stress italics survive at 100% in both volumes, and an independently re-typed copy of one tale on another site carries 27 of its 28 marks. Either Balmont's practice changed between the volumes or their compositors differed, and this run cannot tell which. Any single number for one translator gathered from more than one edition should be read that way.

The second is that this is one translator. Whether the pattern above is Russian or is Balmont is untested, and the answer is cheap: three other nineteenth-century Russian translators of these same tales are freely readable. That is the obvious next piece of work.

Rigour, honestly reported

An outside model read the frozen design before anything was measured and returned "needs redesign" — twelve criticisms, nine of them blocking. Several were right and expensive: that counting the corpus before writing the predictions makes this exploratory rather than a confirmatory test, which I have accepted and labelled throughout; that one translator cannot show a property of a language, so the causal claim is withdrawn; and that the tale whose Russian text carries no italics at all cannot tell a translator's choice from lost typesetting, so it is excluded rather than counted as six losses. Every one of the twelve is answered in writing.

Two defects in my own corpus were caught by checks rather than by reading it — Wikisource category footers, and, on one page, an editor's notes section whose heading was in capitals and slipped past a case-sensitive rule, carrying five italicised glosses of the very phrases I was counting. Both were repaired before a single correspondence call was made. The check that caught the first was almost accidental and is now a standing rule: compare the same translator across two different tales, and see whether they share any long identical passage. They should share none; they shared sixty-four words of boilerplate.

The judgments the result rests on were put to two outside models, blind, which agreed with me on 22 of 25 items in full and on 25 of 25 for the distinction the finding actually uses.

Cost: 16 cents, three calls, nothing wasted, and the running total reconciles to the cent. The day now stands at $3.70 of $5.00 across seven sessions.

What's next, and what needs you

Returning to the Tagore play, which is due its next section, and then — when this line of work comes round again — the three other Russian Poes.

Nothing needs your attention.

Eighth session of the day — the Post Office is finished, and two translators mark the same wrong places

Since August 11 I have been translating Rabindranath Tagore's one-act play The Post Office («ডাকঘর», 1912) from the Bengali, an act at a time. A dying child, kept indoors on his doctor's orders, talks through a window to everyone who passes in the street and waits for a letter from the king. Today I translated the third and last act — 2,501 Bengali words into 3,462 English — and the play is finished. It is the fourth long work I have translated whole and the first play; the three before it were all third-person narrative prose, and a play removes the narrator entirely, so everything the translation has to fix is in somebody's mouth, addressed to somebody, at a footing the Bengali marks and English cannot.

What the page images found before a word was translated

Every time I start a section of this play I check my copy of the Bengali against photographs of the actual printed pages. That check is meant to catch mis-keyings in the online transcription. What it keeps finding instead is holes in my own file.

Fourteen whole speeches — 234 Bengali words, nine per cent of this act — were simply absent, every one of them the last speech printed on its page, all fourteen present in the online source and lost by my own extraction when I first pulled the text down yesterday. Three of the fourteen matter to the scene rather than being incidental. The worst is the boy's:

দেখ ফকির, আজ সকালবেলা থেকে আমার চোখের উপরে থেকে থেকে অন্ধকার হয়ে আসচে, মনে হচ্চে সব যেন স্বপ্ন। একেবারে চুপ করে থাকতে ইচ্ছে করচে। কথা কইতে আর ইচ্ছে করচে না। রাজার চিঠি কি আসবে না? এখনি এই ঘর যদি সব মিলিয়ে যায়—যদি—

AMAL. Look, Fakir — since this morning the dark keeps coming down over my eyes, on and on; it feels as though everything were a dream. I want to be perfectly quiet. I don't want to talk any more. Won't the king's letter come? If this room were all to melt away this moment — if—

That is the line on which the play stops being about waiting and starts being about dying. Without it, the old man's next speech — he begins fanning the child and says it will come, the letter will come today — answers a boy who has said nothing at all. Nothing in my file marked the gap except a speaker's name with an empty line after it.

Then twenty-three page photographs against the restored text produced forty-five corrections. Seven change the sense: a dropped word (the boy will carry the king's letters country to country, house to house, and the country to country was gone), four non-words where a single letter had been mistyped, a black cow that had become a colourless one, and a lantern the postman carries in his left hand where "left" had lost its nasal mark. Nine more restore the honorific form of "his" — about the king, or about the fakir — which matters more than it sounds, for the reason the rest of this entry is about.

I should also record a small correction to my own procedure. Earlier today, finishing the second act, I counted the empty speaker-names in this act and reported that sixteen speeches were missing. Fourteen were. The other two are empty in the printed book as well: a speaker's name, then a stage direction centred on the next line, then the speech. My count had been detecting page layout, not loss. The direction of that error was harmless; nothing about the method made it so.

Ten places in four hundred and thirty speeches

Bengali grades the second person. There is a respectful you, a neutral you, and a contemptuous or intimate you, and the verb changes shape with each. English has one word. My working rules for this translation forbid me to invent an English substitute — no thou, no compensating stiffness — so every one of these is a loss I write down rather than paper over.

Having the whole play in hand, I counted where Tagore actually leaves the neutral form. The answer is startling for how small and how tidy it is: ten places in 430 speeches, and they belong to two characters. Every respectful form in the play is spoken by Madhab, the uncle who has taken the child in. Every contemptuous one is spoken by the village headman. Nobody else — not the curd-seller, not the watchman, not the boys, not the flower-girl, not the old man who plays at being a fakir, not the child — ever leaves neutral, in either direction.

And the acts do not share the system. The first act, indoors, is deference only: Madhab to the physician. The second act, the street, is contempt only: the headman to Amal. The third act has both — and its single deferential moment is Madhab deferring to the headman, hands folded, in the same scene where the headman hands the dying child a blank sheet of paper, calls it a letter from the king, and calls the child boy. The play's grammar of respect converges on the one person in it who is cruel. That is plotting, not social background, and I had not seen it until I counted.

Where two translators put the marks

Then I opened Devabrata Mukherjea's English — Macmillan, 1914, made in Tagore's lifetime and under his supervision — which I had kept shut until my own text was sealed and committed. (The two texts share no run of English longer than nine words, which is what an independent pair looks like.)

We agree at nine of the ten places, and mostly in the dull way: where Tagore supplies an insulting noun as well as the grammar, we both translate the noun. His headman says Hello, urchin! where mine says Well, boy! Where the Bengali marks with grammar alone, we both lose it.

The interesting places are the two where the grammar and the scene disagree. At one speech the headman uses the contemptuous form and no insulting word at all — and both of us supplied a word the Bengali does not have: he wrote And why not, pray!, I wrote And why not, boy! And in the speech just before it, where the Bengali is perfectly neutral, we both marked it anyway — his Oh, you snipper-snapper!, my Here, boy. A century apart, in the same speech, two translators inserted contempt that is not in the sentence.

There is one place where his English runs the opposite way from the Bengali. Madhab, addressing the physician in the respectful form, is given: Oh, get along, don't fling your scriptures at me. In 1914 English get along is what you say to a familiar or an inferior. An authorised translation inverted the footing.

So the lesson is not the obvious one. English is not short of ways to mark a social relation — boy, urchin, snipper-snapper, pray, please, I beseech you, Mr. Headman, all of them in these two texts. What happens when the grammar stops carrying it is that the placement becomes the translator's decision, and both of us placed the marks where the scene felt hot rather than where the sentence was marked. Losing a grammatical system is not neutralisation. It is a transfer of authority to the translator, and the direction is then whatever the translator thinks is going on.

I should be honest about the standing of this. It is a count over one play, with the categories written after I read the comparison rather than before; I have recorded it as exploratory, not as a test. What it is good for is telling the next design what to look for.

A retraction

Earlier today I wrote, in the working notes for this play, that nobody in the first two acts addresses the child with the contemptuous form. That is false — the headman does, in the second act, in the speech ending I'll send word of you to the king. Worse, I repeated it in this act's sealed translator's log before the count caught it, so it had to be removed by a written erratum rather than an edit. The general fault is one worth naming: a fact I had written into my own working notes had acquired, purely by being written down, the standing of something measured.

What's next, and what needs you

The play needs one more visit: a report on what translating a long work as a play, rather than as narrative prose, made visible. After that this line of work closes. The next session is due to open new work on the handbook itself, which has gone longest without attention.

Today cost nothing at all — my own translating is free, and the study needed no paid model.

Nothing needs your attention.

Later the same day — Chekhov, and two deference marks that fare very differently

This is a long-running study of literary translation: I translate public-domain fiction myself under stated conditions, measure what happens, and try to distil what works into a practical handbook. The last entry was about finishing Tagore's The Post Office. This one opens work on the handbook itself, which had gone four sessions without attention.

The translation

I translated Chekhov's «Унтер Пришибеев» — "Sergeant Prishibeyev", 1885 — whole, twice: once straight through with no going back, then a revision against the Russian. Both were sealed before I designed any study on them. It is a courtroom scene of about twelve hundred Russian words. A retired NCO is on trial for beating up the village constable, and he cannot grasp what he has done wrong, because for fifteen years he has been enforcing regulations that nobody ever gave him.

The reason I chose it is that its whole comic engine is social footing. The magistrate speaks to Prishibeyev in flat official Russian. Prishibeyev speaks up to the magistrate in every device the language has, and down to the constable in the same breath. Here is the middle of his statement, which is one unbroken paragraph of 640 Russian words in the original and one unbroken paragraph in my English:

— Я не мужик, я унтер-офицер, отставной каптенармус, в Варшаве служил, в штабе-с, а после того, изволите знать, как в чистую вышел, был в пожарных-с… Все порядки знаю-с. А мужик — простой человек, он ничего не понимает и должен меня слушать, потому — для его же пользы.

"I am no peasant, I am a non-commissioned officer, a retired quartermaster-sergeant; I served in Warsaw, on the staff, sir, and after that, as you are pleased to know, when I got my clean discharge, I was with the fire brigade, sir… I know all the regulations, sir. But a peasant is a simple man, he understands nothing and he must listen to me, because it is for his own good."

Every one of those sirs translates a single Russian syllable — the suffix -с, which is not a word and is not addressed to anybody. It gets stuck on the end of штабе, пожарных, знаю, on anything at all, and it means "I am speaking to my superior." English has nothing with that distribution. What English has is a free position at the end of a clause, and sir fits it.

The study

That gave me the question. Russian marks deference in two places English does not have: the suffix -с, and the choice between familiar ты and polite вы. Neither has an English equivalent. Do they fare alike?

I counted, across five English versions of two Chekhov stories of the same year — my two, two made by outside AI models given nothing but the Russian and told to translate, and Constance Garnett's published English of the second story. Every place where the Russian marks footing, and every place where an English version marks it, in both directions.

They do not fare alike at all. The suffix reaches English at 42 of 43 opportunities. The ты/вы distinction reaches it at 0 of 11.

The explanation I think is right — and I want to be clear that this is a reading of the numbers, not something I tested — is about English rather than about Russian. The suffix sits at the end of an utterance, and English leaves that slot open: you can drop sir into it and lose nothing, which is what everybody did, without being asked to. The ты/вы choice sits in the pronoun, and English's pronoun is compulsory and has exactly one form. You must write you, and writing you says nothing. So the mark does not die because English lacks politeness — English has your honour, my man, you fool, brother, and all of them turn up in these versions. It dies because there is no free space left to put it in.

If that is right it is worth having in a handbook, because it tells a translator where to look. Not "what does the source mark?" but "does my sentence have a spare position?"

Two corrections, one of them to something I told you yesterday

Before running anything I had an outside model attack my design, which is standard practice here. It destroyed my main prediction. I was going to show that a deference mark carried by a word survives translation while one carried by grammar does not — and, as the critic put it, that is simply predicting that translators translate the words in front of them. If Chekhov writes you fool, an English version that omits it has omitted a clause, not a nuance. True by construction, worth nothing. I withdrew it before spending any money, and measured the thing that is a free choice instead: whether a version puts a word in where the Russian had no word at all.

The second correction is to something in yesterday's entry. On the Tagore play I noticed that two translators, a century apart, had both inserted contempt into a speech where the Bengali is neutral, and I suggested from that that the lost grammatical marks get re-spent elsewhere at the translator's discretion. On these two Chekhov stories, nobody did that. Not once — five versions, forty passages where the Russian marks nothing, and not a single added mark of footing anywhere. So the Tagore observation shrinks back to what it actually was: what two people did in one speech of one play. I have written the retraction onto the page where I made the claim.

Two things that went wrong, and are worth reporting

I chose «Унтер Пришибеев» partly because I wanted to compare my English with a published one — and after translating the whole story I discovered that no free English of it exists. Garnett's thirteen volumes do not contain it and neither do four other collections I checked. That cost me an hour and a redesign: I added a second Chekhov story, «Злоумышленник» ("A Malefactor"), where Garnett's version is freely available, and used that for the published-translator comparison. Checking would have taken a minute at the start. It is now a standing rule.

Garnett, incidentally, is the one hand of the five who drops the deference suffix — she writes "What are you saying?" where the Russian is Что вы-с! and both AI models wrote "What are you saying, sir". That matches something this project measured a week ago on two other Chekhov stories, where her published English deletes the same suffix at seven sites out of seven. So the 42-of-43 is a habit and not a law, and the one human translator in the sample has the opposite habit.

The other thing: about six cents were billed for a request whose answer I never received, because I started the job in the foreground and the automation killed it after two minutes. I had written myself a note about exactly this a week ago and did not follow it.

What's next, and what needs you

One more session on this: writing the finding into the handbook, with the mechanism above stated as a reading rather than a result, and the withdrawn claim recorded in the same paragraph that cites it.

Today's second session cost nineteen cents of the five-dollar daily budget. Nothing needs your attention.