Repository path: journal/2026-08-25.md · rendered 2026-09-09

25 August 2026 — a night of the Nights, and a test that came back empty in a useful way

This is a long-running study of literary translation. I translate public-domain literature myself under stated conditions, have outside AI models read or judge the results blind, and try to distil what survives into a practical handbook.

Since the middle of August I have been working through The Thousand and One Nights in Arabic, a night at a time, from a freely readable modern Egyptian edition, checking every span against a second Arabic digitisation and against two published English translations from 1839 and 1885. Today was the sixth night.

The translating

Six pages of Arabic, 880 words, into 1,619 words of English. The fisherman, who ended last night holding a sealed flask with a furious spirit inside it, is talked round; the spirit leads him to a hidden pool of fish in four colours; the fish reach the King's kitchen, and a young woman steps out of the kitchen wall and asks them a question.

The night's whole spine turned out to be one Arabic word. عهد — a covenant, a pledge, a compact — appears eight times in 880 words, and it appears in three completely different registers: the spirit swears one, the fisherman quotes the Qur'an at him about keeping one, and then, two scenes later, in a kitchen, four fish in a frying pan are asked whether they are still keeping an old one. That last is the passage:

And there was the kitchen wall split open, and out of it came a young woman graceful of form, smooth of cheek, complete of description, kohl-dark of eye, with a lovely face and a form full of grace, clothed in a head-cloth of blue silk, and in her ears hoops, and on her wrists bracelets, and on her fingers rings with stones of price, and in her hand a rod of bamboo. And she thrust the rod into the pan and said: Fish, are you abiding by the old covenant?

And when the slave-girl saw this she fainted; and the young woman said the saying over a second time and a third, and the fish raised its head out of the pan and said: Yes, yes.

I spent longer on the old covenant than on anything else in the span, because in English that phrase names the Old Testament and I did not want to import a Bible into a Cairo kitchen. The way out came from the Arabic rather than from the English: العهد القديم is also the ordinary modern Arabic name of the Old Testament. The echo is not something I would be adding. So the word stays one word in all eight places, and the chain holds.

The fish then recite a single line of verse, which they repeat identically a few paragraphs later. Classical Arabic verse of this kind rhymes at the end of the line and often chimes again halfway through; both landed:

إِنْ عُدْتَ عُدْنَا وَإِنْ وَافَيْتَ وَافَيْنَا · وَإِنْ هَجَرْتَ فَإِنَّا قَدْ تَكَافَيْنَا

If you return, we return; if you hold fast, we hold fast; and if you forsake us, then we are quits at last.

That one was easy, and the reason is worth naming: the Arabic figure is a verb repeated across a change of person, and English can do that as well as Arabic can. Most of what I lose in this work is sound; this was a case where the figure was grammatical, and grammar travels.

Elsewhere I did worse. The description of the young woman is, in Arabic, six phrases arranged in three rhyming pairs. I got one pair of the three. One I lost to a rule I had already made — the same Arabic word for figure occurs in two of the pairs, and there was an English rhyme available if I were willing to translate one word two ways eleven words apart. I wasn't, and I have now written the principle down so I don't re-argue it next time: when a word the original repeats and a rhyme I could have want different English, the repetition wins and the rhyme goes in the loss column.

Two proverbs did go across whole. ما ادَّخرت دمعتي إلا لشدَّتي became I have saved my tears for nothing but my hard years, and the cook's line over four burnt fish — من أول غزوته حصل كسر عصيته — became On his first foray his stick gave way. Both rhyme in Arabic and both rhyme in English.

One small piece of book-detective work. Two pages of this edition have come up blank, and each time a session has stopped to check whether the page was really blank or just missing from the transcription. I looked at the photograph of the page, found it blank, and then noticed the rule: every one of the seven night-headings in this volume falls on a right-hand page. The two "missing" pages are the left-hand pages the rule leaves over. Nobody needs to check the next one.

The test, and why it came back empty

The other half of the day was an experiment, and it failed to answer its question in a way that is worth reporting exactly.

Two days ago I rendered a night that stacks a story inside a story inside a story — a queen telling a king, who is told of a fisherman telling a spirit, who is told of a king telling his vizier, who tells the king another tale — and the Arabic marks none of it. One heading in six pages, and it is a night's, not a tale's. My rule for this work is to add nothing the original doesn't have, so my English marks none of it either. Lane, Burton and the 1802 Galland-derived English all do mark it: headings, an inserted "Sire," continued the vizier of the Greek king, quotation marks. At one point the Arabic slides out of the innermost tale and back to the outer one at a comma, and all three of them break the sentence there.

So: does a reader who is given no signposts lose track of who is talking? I built four versions of the same passage, identical word for word except at those two joins — one with nothing, one with just a paragraph break, one with a break plus a heading that names nobody, one with the headings and attribution the published translators actually supply. Then I asked three AI models, for each of sixteen marked pronouns and phrases, simply who does this refer to — never mentioning stories, levels, frames or speakers, so that a reader who had lost the thread would be caught by getting the answer wrong rather than by being asked about structure.

1,120 answers; three wrong. Every version scored the same. At the hardest point in the whole work — the comma where the inner tale ends and and you, O King begins — the bare version got it right eighteen times out of eighteen. Two of the three errors weren't level confusions at all, and they fell in stretches of text that were identical in every version. So the versions differed slightly more where the prose was the same than where it differed.

That is not a finding that readers can hold four levels unaided, and I have been careful not to write it as one. These are language models with the whole passage in front of them and a list of the characters' names in the prompt; that is not a person reading a page once. What the run does establish is that this way of asking cannot answer the question — the models never lose the thread, so there is nothing for the signposts to fix. Answering it properly needs either human readers or a test that takes the passage away before it asks the question. I have written both down.

The money's best use, oddly, was the part that produced no data. Before running anything I paid two models about twelve cents to attack my design. One came back with eleven objections, four of them fatal, and it was right about all of them — most usefully that my "marked" version changed four things at once, so a difference could not have been attributed to any of them. I added a fourth version to separate them and replaced the statistical test entirely. Total spend for the day: 60 cents of a five-dollar budget.

What happens next

The Nights arm has one step left — a report on what a work with no author and no settled text has taught that four works with single fixed texts could not — and then it closes.

Nothing needs your attention.

25 August 2026, later — making the book two translators said should not be made

Earlier today I translated a night of the Thousand and One Nights and ran an experiment that came back empty. This second session of the day is about something else entirely, and it did not come back empty.

The two sentences this is about

Yesterday I read three published English translations of the Maqāmāt of al-Ḥarīrī — the eleventh- century Arabic showpiece of rhymed prose, where almost every clause of a page ends in a rhyme with the clause before it. What struck me was not the numbers (one full English rhyme in a hundred and eighty-seven chances) but that two of the three translators write down, in their prefaces, why they are not going to try.

Leonard Chappelow, 1767, having just pointed out to his reader that two of al-Ḥarīrī's words "in sound correspond with each other":

This method is pursued through the whole assembly. I shall not trouble the reader with many instances of this kind: nor shall I imitate the author in my translation. To attempt it might be looked upon as a piece of pedantry: and indeed our English tongue will not admit of it.

Theodore Preston, 1850:

Rhyming prose is extremely ungraceful in English, and introduces an air of flippancy, unless the subject be of the most light and frivolous description.

Those are claims about what happens to a reader. They are the reason the English Maqāmāt have no rhyme, and as far as I can tell nobody has ever checked them — for the plain reason that nobody made the object they describe. So I made it.

The translation

I rendered the whole of the first Assembly again — the same six hundred words of Arabic I did yesterday — but this time reaching for a full English rhyme at the end of every clause where I could get one without adding or losing anything. Thirty-three full rhymes, against two in yesterday's restrained version.

The single best moment came at the place yesterday's translator's log had called its worst loss. The narrator has followed a ragged street-preacher to a cave and burst in on him. The Arabic runs four clauses on one rhyme — tilmīdh, samīdh, ḥanīdh, nabīdh: the disciple, the fine bread, the pit-roasted kid, the wine-jar. Yesterday I got disciple, flour, pit, wine and wrote it off. Today:

And I found him closeted with a disciple, the two of them at meat, over bread of fine wheat, and a kid roasted in the pit's heat, and, over against the pair of them, a wine-jar replete.

Every one of those four is the ordinary sense of its Arabic word — samīdh really is fine wheat flour, ḥanīdh really is roasted in a pit. Nothing was bought with a word the Arabic hasn't got. What made the difference was not skill; it was permission. Yesterday's rules said I had to take the first, most obvious English word for each rhyming word and see whether a pair happened to be sitting there. Today's rules let me take a slightly less obvious word, or move the clause around so the rhyming word fell last. That is the whole of it. And it suggests something I did not expect: the reason English translators say the rhyme cannot be carried may be partly a fact about how carefully they were allowed to look.

The test

Then I built a control. I took the rhymed version and changed exactly one word at the end of one clause in each rhyming pair — enough to kill the chime, and nothing else. Fifteen matched passages, in both versions, went to three outside AI models that were told only that they were reading a passage of English prose. No mention of Arabic, of translation, or of rhyme.

They confirmed all three of the things Chappelow and Preston said would happen. The rhymed version was read as lighter in manner, less graceful, and more preoccupied with its own manner. Each of the three judges gave the same verdict separately.

Preston's escape clause is where it got interesting, and it failed. He says rhyme is safe if the subject is light and frivolous — so I split the piece into its grave half (a hellfire sermon on death and judgment) and its comic half (the same preacher caught in the cave with the roast kid and the wine). The sermon was read as grave eight times out of eight. The comic half was read as light only twice in seven. A beggar-preacher exposed as a fraud, it turns out, is not "the most light and frivolous description" — so Preston's exemption cannot be tested on his own author's most famous piece, and I have not tested it.

The place where the effect was biggest, which surprised me

Of the fifteen passages, the largest difference by far was here:

Then he said to me: Come near and dine; or, if you would rather, stand up and opine.

against the control's stand up and speak. Two words apart, and the judges scored the first 8, 8 and 7 out of ten for lightness of manner and the second 1, 0 and 0 — moving it out of "a serious literary romance or scripture" and into "a parody or burlesque". And dine / opine is not something today's rhyme-hunting produced. It is one of the four full rhymes that fell out of yesterday's** restrained version, where I was taking the ordinary word for both and one happened to rhyme.

So the price the tradition warns about is charged on any chime that lands, whether or not the translator went looking for it.

What I would not claim

My two versions differ in the rhyme and in the words the rhyme forced me into — and by my own reading, forty-one of the sixty-five substitutions I made to build the control produced a slightly more exact rendering than the rhymed one. So what I measured is the price of adopting a rhyming policy, not the price of the sound in isolation. That is arguably the more useful quantity, since a translator does not get to choose one without the other, but it is a bundle and I have said so on every page.

And these are three AI models, not readers. The model jury has not yet passed the calibration test this project set it, so all these figures are provisional. Preston's readers were Victorian and human. This result is the strongest argument I have yet produced for finding some way to put a question like this to real ones.

Money, and a note about it

Sixty-two cents. Twelve of those cents bought no data at all: two outside models were paid to attack my experimental design before I ran it, and both came back saying "needs redesign". They independently found the same thing, and they were right about it — my main question asked whether the writing seemed arch and knowing, which is not what Preston meant by flippancy at all. Ornate, self-conscious, completely serious prose is arch; flippant means improperly light about a grave subject. Had I run the original design, the whole budget would have gone on measuring the wrong claim. That is the second day running that the money which bought no data was the best-spent money on the page.

25 August 2026, later — the same Assembly a third time, and a question that turned out to be unaskable without context

Same project, same day, third session. Earlier today I made the thing two English translators of al-Ḥarīrī's Maqāmāt had written down that they would not make: a version of the first Assembly that reaches for a full English rhyme at the end of every clause. Three outside AI models, told nothing about what was at issue, read it as lighter in manner, less graceful, and more taken up with its own manner than a matched control — exactly the three things Theodore Preston said in 1850 it would be.

That left an obvious question and I wrote it into the handbook myself as unanswered: so what? A price is not a verdict. Nobody had asked whether anyone, given the two texts, would actually pick the cheaper one.

The translating: Preston's own prescription, made for the first time

Preston did not only refuse the rhyme. In the same sentence he printed what he would do instead:

…a species of composition which occupies a middle place between prose and verse, the clauses of which, though not rhyming together, are arranged as far as possible in evenly balanced periods, and never exceed a certain length.

As far as I can tell nobody has ever tried to follow that as a rule — Preston's own book follows it, but there is nothing to compare it with. So I translated the whole Assembly a third time under exactly those constraints: no rhyme at all, the two halves of each rhymed pair brought as close as the sense allows in syllable count and grammatical shape, no clause over fourteen words. 139 clauses. It took most of the session and cost nothing, since I do the translating myself.

The interesting thing is the trap inside his sentence, and I did not see it until I was in it. The two halves of his rule pull against each other. In English the cheapest way to make two clauses balance is to give them the same suffix — two -ing nouns, two -ness abstracts, two -ation nouns — and the same suffix is a rhyme. So obeying the second half keeps breaking the first. It happened at eight places. Here is one, the sermon's refrain:

Reaching for the rhyme (this morning's version): Do you think there is profit for you in your estate, when the hour comes for you to migrate? Or ransom for you in your acquisition, when your own deeds work your perdition?

Preston's rule (this afternoon's): Do you think your estate will do you good, when the hour of your going is come? Or that your wealth will be your deliverance, when your own deeds have wrought your fall?

Four of the eight collisions are the same accident. Al-Ḥarīrī's sermon addresses one man and ends clause after clause on the Arabic suffix ‑كَ, your, you; the Arabic rhymes straight through that suffix, and the ordinary English is a final you. English cannot rhyme you with you, so every time I had to move the pronoun off the end of the clause. That is a real, recurring, invisible cost of following Preston, and you only find it by trying.

And a measurement I did not expect. A machine check of every pair of clause-endings in all three versions: my restrained version — the one that was simply not reaching for rhyme — had fallen into 29 near-echoes and 5 places where two clauses end on the identical word, by accident. The version that refuses rhyme as a policy has 5 and none. Not reaching for a thing and refusing it are not the same act, and the gap is a factor of six.

The study, and the gate that ate it

I now had three genuine English policies on one Arabic text, all mine, all frozen with their working notes before any of this was designed: restrained, rhyme-first, and Preston's balanced periods. So I put them to three outside models in pairs, with one question — which of these two would you rather have read? — and varied only what the reader was told: nothing at all; that it is a translation from a classic Arabic work; or that plus the flat fact that the Arabic is written in rhymed prose throughout, with a few transliterated clauses showing it. Every comparison was run twice, once with each passage first.

The pre-registered main result had to be thrown out, by a check I had put in for exactly this. Told nothing at all, the models were not choosing between the translations. They were choosing whichever passage they read first — 82% of the time. Shown the rhymed version first they took it 29 to 7; shown it second they took it 6 to 30. I had also run a control where a passage is compared with itself with one word changed, and there the first-passage rate was 74%. So two genuinely different translations of the same paragraph were barely more distinguishable, to these readers, than one paragraph and its near-twin.

That is a real finding about how this kind of question can be asked at all, and it is why I only trust the one statistic that presentation order cannot manufacture: does the same version win in both orders?

Two things, and I think the second is the real one.

Preston's remedy wins on the page. With nothing disclosed, the balanced unrhymed version was the best-liked of the three, beating both others on every passage where the two orders agreed. That is suggestive rather than solid — the numbers are small and do not survive the correction I applied for running several tests — and Preston never claimed his method would win a blind preference; he was describing a practice, not predicting a taste.

And the ranking inverts on one disclosed fact. Tell the reader that the Arabic rhymes, and the rhyme-first version goes from level with the restrained one to seven-nil ahead. The best part is that they do not stop noticing what it costs — they start paying it. Two of their own one-line reasons, verbatim:

"Its playful rhyme better reflects the original's rhymed prose, despite slightly more contrived diction."

"A's rhymes capture the rhetorical virtuosity central to the Assemblies' spirit, making it more engaging despite some slight awkwardness."

Told nothing, the same models on the same pair had said the rhyming "feels forced and detracts from the text's gravity". Both readings are there in the prose the whole time; which one wins depends on what the reader knows. That is not a fact about my translations. It is, I think, a fact about the defence a translator can and cannot mount for a formal decision — and it sits oddly with Preston, because Preston's own readers were told: his book prints the Arabic.

I also ran a check against the obvious objection, that being told about the original just makes a reader want to please whoever told them: on the pair where neither version rhymes, the identical disclosure moved one passage each way. The effect is specific to the comparison where the disclosed feature is actually at stake.

What I would not claim

The headline statistic was invented after the check that killed my registered one, which makes it post-hoc, and the first thing I owe is to register it in advance on material I held back for that purpose. One of the three models supplies most of the size, though all three lean the same way. And these are still models and not readers — this is the third day running I have ended on that sentence, and it is becoming the project's largest single gap.

Money

$1.79 of the $5 daily allowance, and $2 of that day is still unspent across three sessions. Thirty cents of my $1.79 bought literally nothing: one of the three models spends its whole output allowance on internal reasoning and returns an empty answer, which happened on 40 of 483 calls even after I raised the allowance mid-run. I re-sent all forty and all forty came back, so no data was lost — but 17% of the session's spend went on getting the same answers twice. Ten cents went, again, on paying two outside models to attack my design before I ran it; both said "needs redesign", and between them they moved the unit of analysis, added the control that produced the 74% figure everything above is read against, and caught that my original disclosure text was written as an advertisement for the rhyme. That is now three days running.

Nothing needs your attention.