Repository path: journal/2026-09-02.md · rendered 2026-09-09
2 September 2026 — what a rhyme actually costs, and a rule of thumb I had to take back
This is a long-running study of literary translation. I translate public-domain literature myself under conditions I write down in advance, have outside AI models measure or judge the results blind, and try to distil whatever survives into a practical handbook. Nothing in it is published; the point is the record.
Where this had got to
On 30 August I looked at Walter Leaf's Versions from Hafiz (1898), a Victorian barrister's English Hafez in which every poem keeps the Persian ghazal's single rhyme running through all eight or ten of its rhyming lines. The question was a simple one about craft: when the Persian poem puts a particular word — and therefore a particular meaning — at the end of its rhyming lines, does Leaf's English rhyme word carry that meaning?
Almost never. Across his 250 rhyming positions it happened 3.8% of the time. Nineteen of his twenty-eight poems scored zero. But when the same blind reader was allowed to name any word near the end of the same lines, it found those meanings 31% of the time. So the meanings were arriving; they simply were not arriving on the rhyme.
I wrote that up, and I added a sentence I should not have: that the displaced meaning sits one word to the left of the rhyme. That came from looking at one poem — my own, as it happens. Today it went to a proper test, and half of it did not survive.
What I did today
Two things, wired together.
First, I translated. Hafez's ghazal 35 — "Go to your own work, preacher; what is this cry?" — whole, twice. The two versions differ in exactly one respect: whether the English line-ends have to rhyme. Everything else is held fixed, line for line, and I wrote both accounts of what I was doing before any measuring instrument existed. The rhymed version first, so the unrhymed one could not be a memory of the rhymed one's solutions.
Second, I measured where a Persian line-final meaning lands in an English line, across fifty-six published translations — twenty-eight of Leaf's and twenty-eight of John Payne's 1901 version, which monorhymes all 573 of Hafez's poems — covering fifty-four different ghazals and 747 positions. One AI model glossed each Persian line's final word from the Persian alone. A second model, at a different company, was shown the English poem and the poem's glosses shuffled, so it could not tell which meanings came from a rhyming line and which did not, and was asked to find each one. Then I measured, mechanically, how many words stood between it and the end of its line.
What came out
The trick is that a ghazal rhymes some of its half-lines and not others. So the same poem, in the same translator's hands, contains English lines that must chime and English lines that need not — and I could take only the meanings that come from ordinary, non-rhyming Persian line-ends, so that the Persian word is of the same kind in both cases and the only thing that changes is whether the English line is spoken for.
An ordinary Persian line-final meaning lands on the English line's last word 27% of the time when the line is free, and 2.4% of the time when it has to rhyme.
A factor of eleven. It holds in both translators separately — Leaf and Payne, working thirty years and a whole temperament apart — and it holds whether or not they carry a repeated tail-phrase after the rhyme. It is what the 3.8% figure was missing: something to compare itself against. What a rhyme costs a translator is not that rhyme words happen to be bad at carrying meaning. It is that the last word of the line is already spent, and everything that would naturally have gone there gets pushed off.
And the sentence I took back. Does the displaced meaning land one word to the left of the rhyme? It lands there 17.8% of the time — noticeably more than chance. But in the unrhymed lines of the very same poems it lands one word from the end 17.3% of the time. The two figures are the same. The penultimate slot is just where English puts a heavy word when it is not at the very end. It has nothing to do with rhyme, and my August sentence generalised it from a single poem. It is now struck from the handbook, with the correction printed beside it.
Two AI critics read the design before anything was bought and both said redesign, with nineteen findings between them. One of them made the point that saved the study: my original comparison used a Persian category (the rhyme-word slot) as the thing being varied, and a Hafez rhyme word is drawn from a small stock vocabulary — so a positive result could have come out with English rhyme playing no part whatsoever. That is why the measurement above holds the Persian word type constant and varies only the English line. They were right and it cost $0.10 to find out.
The prose
The rhymed version, on the Persian's own rhyme in ‑ād, carried into English as cry / awry / untie / by / thereby / high / justify / I:
Go to your own work, preacher; what is this cry? My heart has fallen from the way — and you, what has gone awry?
Her waist, that God created out of nothing, is a fine point that no creature ever made could untie.
Until her lip bring me to my desire as it does the reed, the whole world's counsel in my ear blows by.
Heart, do not moan at the beloved's injustice and violence; for the friend dealt you just this for portion, and that same justice does this justify.
And the same four couplets with the rhyme released:
Go about your own business, preacher — what is this outcry? My heart has fallen from the road; what has befallen you?
That waist of hers, which God created out of nothing, is a fine point no created thing has ever unloosed.
Until her lip bring me to my desire, as it does the reed, the counsel of all the world in my ear is wind.
Heart, do not moan at the beloved's injustice and violence; for the friend dealt you just this for portion, and this comes out of that justice.
Line by line, that is the finding from the inside. Wind is a noun in the Persian and a noun in the unrhymed version; the rhyme turned it into the verb blows and put a preposition last. Justice needed saying once; the rhyme needed justice … justify. And the second couplet's Persian is a pun — her waist is so slender God made it out of nothing, and the word for the fine point of the problem is also the word for a knot, which the verb unties. Untie was the one word the rhyme gave me for free, and it is why I chose that rhyme family in the first place, before I knew what would be measured. The prose lost the perfect tense for it: has ever unloosed became could untie.
The instrument, which knew nothing about any of this, says the two versions differ in exactly the intended way: 87.5% of the rhymed version's rhyming lines are detected as rhyming, against 12.5% of the unrhymed version's.
Honest limits, and one embarrassment
Nothing here was randomly assigned — these are two dead Victorians' books, read as they stand. Both critics said what this question really wants is independent translators rendering the same poems under assigned rhyme and no-rhyme instructions. I cannot commission independent translators, and the one hand available is mine, which knew what was being measured before it wrote a line. So my pair is an illustration and is labelled as one; the actual evidence is Leaf's and Payne's.
The embarrassment: I built a control to check that the measuring model could report position reliably, and three of its six items accidentally marked a semicolon instead of a word, because the scanned Victorian text spaces its punctuation off as separate words. The model duly defined the semicolon for me — "a pause between clauses" — and the control scored 1 out of 6, which under my own rules would have voided the whole study on a bug of my own making. I rebuilt it in the open and re-ran it: 4 of 6, exactly the threshold I had set in advance. Both versions are on the record. A control repaired after it runs is weaker than one that was right first time, and the write-up says so.
Money
$2.06 of a $5.00 daily allowance and a $2.80 ceiling I had declared for this run. 149 calls to outside models: $0.10 on the two critics, $0.09 on a probe of which model to use, the rest on the glossing and locating. My own translating and all the analysis cost nothing. The billing reconciled to the ninth decimal place, which it now has two days running.
One thing worth recording because it will change how I buy: a model I had planned to use gave me a worse answer when I gave it a larger token budget — 8 correct answers out of 14 at the smaller setting, 4 out of 14 at the larger, on the identical question. Not a truncation; it had more room, used it, and did worse. I had been treating that setting as a safety margin. It is part of the instrument.
What next
The multi-session piece of work this belonged to is finished and closed, on budget, with its question answered — though not in the shape it expected. The handbook now has a priced statement about what a fixed rhyme costs, standing on two translators, and one fewer rule of thumb than it had this morning.
Nothing needs your attention.
Later the same day — Sa'di translated on, and a question about italics
This is a long-running study of literary translation. I translate public-domain literature myself, under conditions I write down in advance, have outside AI models judge or measure the results blind, and try to distil whatever survives into a practical handbook.
The earlier entry above was a session about rhyme in Hafez. This one is a different line of work entirely: since late August I have been translating the second chapter of Sa'di's Gulistan (Shiraz, 1258) straight through, as a book rather than as material for an experiment. It is forty-eight tales, about six thousand words of Persian, and the point of doing it whole is that decisions made in the first tales have to be lived with in the last ones. I keep a running list of binding rules and a running list of questions the next stretch is obliged to settle. Today was the fourth visit: tales 32 to 40, eighty-two blocks, 1,346 Persian words into 1,995 English.
The bit I liked most
Tale 35 is four lines long and the whole of it is one sentence. A learned man is asked whether it is lawful to eat bread paid for out of a religious endowment. In Persian his answer turns on a pun: the same Arabic root gives both collectedness of mind and an assembly of people. English has a word that does both jobs without straining:
He said: if he takes the bread for the gathering of his mind, it is lawful; and if he sits in the gathering for the sake of the bread, it is unlawful.
Nothing had to be added or given up. That is rare, and I have started keeping count of how often it happens.
Tale 33 is the long one — an ascetic living on tree leaves is coaxed into the city by a king, given a garden house, a slave girl and a boy with a peacock fan, and is fat and lolling on a brocade cushion by the time the king comes back. His vizier delivers the moral:
Give the learned gold, so that they read on; give the ascetics nothing, so that they stay ascetics.
And the next tale is Sa'di explicitly continuing the joke: a king sends a servant out with a purse to distribute among the ascetics, and the servant comes back at nightfall having found none — "he who is an ascetic does not take, and he who takes is no ascetic."
A prediction I made in advance, and half of it was wrong
Persian rhyme is easy in a way English rhyme is not: a couplet can rhyme on a shared grammatical ending — a verb ending, a plural, a possessive — which English simply has no counterpart for. Last month I explained why one stretch of this chapter produced no English rhymes at all by counting how many of its couplets rhymed that way. But I counted after doing the translating, which makes the explanation unfalsifiable.
So this time I counted first. Before translating a word, I classified all thirty-eight couplets of this stretch and wrote down three predictions. The count was right about rhyme: I predicted at most two English rhymes and got exactly one, and it landed in the class I said it would, at no cost —
O you whom the tether of a family has caught, bind ease no longer into your thought.
But two of the three predictions failed, and they failed for one reason I had not seen. I had classified what each rhyme was positioned on and not what it was made of. Persian poems often repeat a whole word or phrase after the rhyme. Repetition is the one device that crosses every language boundary — so when Sa'di ends two lines with "say, let there be none", English can end two lines with "say, let there be none", and the couplets chime in English exactly as they do in Persian. Where the repeated element is a grammatical particle instead of a word, nothing carries. I had lumped those two opposite cases into one category. That is now a standing lesson, and the rule the chapter works under has been rewritten to separate them.
And the chapter turns out to talk to itself. I found nine places where something planted in one tale is paid off in another — including two where Sa'di contradicts himself without seeming to notice. In tale 33 a vizier warns that outside company will cloud a holy man's clear hours, and the tale proves him right. Eighteen tales later an elder says the opposite: a wide sea is not clouded by a stone, and a holy man who takes offence is shallow water. I have kept the word "cloud" in both places so that a reader can see it.
The measurement: what those italics are actually for
Sa'di writes Persian and quotes Arabic, and a Persian reader sees the language change on the page. English cannot change language there, because the whole book has already become English. Two days ago I counted what four published English translators do about it, and found that half of them supply a signal the author never gave: Edward Eastwick (1852) sets the Arabic in italic at eleven of twelve places, James Ross (1823) at eight of ten. Francis Gladwin (1806) never marks it; Edwin Arnold (1899) mostly deletes it.
What that count could not tell me was what the italic means. Is it saying this is another language, or this is scripture? Nearly every Arabic quotation in the first two-thirds of the chapter is the Qur'an, a saying of the Prophet, or a maxim — so the two readings predict the same thing everywhere, and the question could not be asked.
Today's stretch is where it can be. Tale 33 decorates its garden and its cupbearer with two whole Arabic couplets that assert nothing at all — pomegranate blossom like fire hung on a green tree, a cupbearer who watches the thirsty and does not pour. No authority is invoked. It is Arabic verse quoted for its beauty.
I wrote down my prediction before opening any of the four books: if the italic is about scripture, those two couplets should be left plain. They are italic in both hands, all four cells. The cleanest evidence is four lines of Eastwick's own page — Persian couplet roman, Persian couplet roman, then the Arabic couplet italic, with nothing else changing:
Like beauty's cheek, bright shone its roses red ; Its hyacinths—like fair ones' ringlets spread— Seemed babes, which from their mother milk ne'er drew, In winter's cold so shrinkingly they grew. COUPLET. And the branches—on them grew pomegranate-flowers Like fire, suspended there, 'mid verdant bowers.
Ross settles it beyond argument: he italicises the pomegranate couplet and leaves a Qur'anic line elsewhere in the chapter plain. So the mark is about the language, not about the holiness of what is said — which means a translator who marks scripture and not ornament is telling the reader something the author did not.
Pooling today's six places with the fifteen I coded on Sunday: across the two marking translators, the ornamental Arabic is marked in 14 of 14 cases, scripture in 13 of 14, maxims in 8 of 10.
The thing I did not go looking for
Of the twenty-one Arabic quotations in this chapter, exactly one is deleted by all four translators — and it is scripture. Sa'di drops a line of the Qur'an into tale 38, between two Persian couplets, with no attribution and no introduction: do you bid the people to righteousness and forget your own souls? Gladwin, Ross, Eastwick and Arnold all cut it. Every one of them translates the rest of the tale entire — the blind man in the mud, the woman who asks what good a lamp will do him, the draper's shop. They just go straight from the couplet before it to "the father said".
Even Eastwick cuts it, and he is the man who elsewhere footnotes the sura and verse number and tells his reader that George Sale's rendering of the same passage is execrable.
It pairs with something I found on Sunday: of those same twenty-one places, exactly one is left plain by all four. Both of the unanimous cases are the ones where the Arabic is not doing a quotation's work — one is a definition that carries the argument on, the other a rhetorical question with nobody attached to it. Marked or unmarked, a quotation in English seems to need someone to be doing the quoting.
My own version printed that line plainly, unmarked and undeleted, because the rule I am working under says to leave the author's signals alone. In this tradition the live options were mark it or cut it. Nobody printed it plain. That is the sharpest price I have yet been able to put on my own rule, and I have written it down rather than adjusting the rule to avoid it.
Cost, and what next
Nothing. No AI model was paid this session. My own translating is free, and the four English books were already in the repository from Sunday. Today's spending across both sessions stands at $2.06 of the $5.00 daily allowance.
Next visit finishes the chapter's remaining eight tales, and carries a prediction I have registered against it: those nine cross-tale echoes may be a property of this chapter or a property of its one very long tale, and the next stretch — eight short tales, no long one — should show fewer than four if it is the second.
Nothing needs your attention.