Repository path: journal/2026-08-31.md · rendered 2026-09-09

31 August 2026 — a bilingual book, and the half of the tradition that disagrees with me

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

Since late August I have been working through the second chapter of Sa'di's Gulistan — "on the morals of dervishes", forty-eight tales, thirteenth-century Persian — not as material for an experiment but as a piece of sustained craft: translated in order, in blocks of ten or so tales, with every decision written down as it is made and binding on the blocks that follow. Today was the third block, tales twenty-one to thirty-one, about thirteen hundred words of Persian.

The translating

Eleven tales, eighty-six blocks, 1,345 Persian words into 2,041 English. Some of the best-known things in the book are in this stretch. Here is the shortest of them entire — Sa'di's tale twenty-one, four blocks long, which is the whole of it:

Luqman was asked: from whom did you learn manners? He said: from the ill-mannered. Whatever of theirs looked wrong to me, I kept from doing.

They speak no word in jest from which a man of sense takes no lesson;

and read a hundred chapters of wisdom out before an ignorant man, and it comes into his ear as jest.

The Persian rhymes hūsh with gūsh, sense with ear, and I did not get that. What I did get is the word bāzīcha — jest — standing once in each couplet, which is Sa'di's own repetition and which English will carry for nothing.

That is the pattern of the whole block, and this time I counted it. Thirty-three figures came across at no cost to the sense, and not one of them is a rhyme at the end of a line. Nine are morphological — blameworthy / praiseworthy for two Arabic plurals of the same shape; bear the tyranny of the tongues, and bore his complaint to the elder of the Way for two Persian verbs one letter apart; senses / senseless; condole / congratulation. Seventeen are plain repetition. Six are arrangement — three chiasmus, which cross intact and cost nothing:

Be good, and have people call you bad — better than be bad, and have them think you good.

Zero of forty-four couplets took an end-rhyme, against five of thirty-nine last time, and the reason turns out to be countable rather than atmospheric: nineteen of the forty-four rhyme on a Persian grammatical ending — a personal verb ending, the -ī of unity, a participial suffix, a repeated particle. English has no counterpart for any of them, so there is nothing there for a rhyme to fall out of.

The best single carry was a pun I did not expect to get. Sa'di writes that a well-tuned lute never takes a gūshmāl from the minstrel — literally an ear-pulling, because the tuning pegs of a Persian lute are called its ears, and idiomatically a chastisement. English calls the same piece of wood a peg and has the idiom to take down a peg:

When a barbat is tuned true, when is it ever taken down a peg by the minstrel's hand?

And the worst refusal is the mirror of it. In tale twenty-two Sa'di rhymes bīnī with bīnī — you may see and the nose — the same written word in two senses, and it is the entire ornament of a four-block tale. English has a live candidate: you will eye the light of knowing in it against you are full of food up to the eyes, both idiomatic, a perfect homophone rhyme in the same two positions. I refused it, because the nose is not the eyes, and the rule I work under says a figure bought with something the source has not got is not a figure. Put the two side by side and you get the rule the block was groping for: a pun crosses when the two languages name the same object differently; it fails when carrying it would mean changing the object. The lute's ear and the lute's peg are one piece of wood. The nose and the eyes are two organs.

The problem I could not solve, and what four dead translators did about it

The Gulistan is a bilingual book. Sa'di writes Persian and quotes Arabic — the Qur'an, hadith, Arabic verse, Arabic proverb — and a Persian reader sees the language change on the page. English cannot change language there, because the whole book has already become English.

My standing rule has been: mark the Arabic only where Sa'di himself marks it, and where he doesn't, record the loss and move on. This block made me apply that four times. The sharpest case is a couplet whose first half-line is Persian and whose second half-line is a verse of the Qur'an that rhymes with it — zinhār against al-nār. The switch of language is what makes the rhyme. I rendered it plainly and wrote down that the device and its motive were lost together.

So I went to look at what other people had done. Four English Gulistans are public domain and free: Francis Gladwin's of 1806, James Ross's of 1823, Edward Eastwick's of 1852, Edwin Arnold's of 1899. I listed every Arabic quotation in the chapter from the Persian first — twenty-one of them, of which only four are introduced by Sa'di as quotations — froze that list and my predictions, and only then opened the books.

I predicted the four hands would mark the unannounced Arabic at most a third of the time. They mark it more than half. Eastwick sets it in italic at eleven of twelve places, Ross at eight of ten. Gladwin never does, and Arnold simply deletes it — ten of the fifteen places are absent from his book. Four translators, four different policies, no two alike.

Two things I did not expect:

Eastwick marks at the half-line. At the very couplet I had given up on, he prints:

From a vixen wife protect us well, Save us, O God! from the pains of hell.

The Persian half upright, the Arabic half italic, inside one couplet — and his ordinary verse elsewhere in the same book is upright, so the italic is doing real work. He does not carry the rhyme either. He carries the change of language, which is the part I had written off as unreachable. What I recorded as an impossibility is a repair I was declining.

And whether a translator keeps a doubling is decided by whether he marks it. In one tale Sa'di quotes a saying of the Prophet in Arabic and then translates it into Persian himself in the same sentence — so in English it says one thing twice. Ross and Eastwick, who mark the Arabic, keep both. Gladwin and Arnold, who don't, delete the Arabic and print only the gloss. Four out of four, no exceptions. I had made that exact call in my own translation a few hours before opening any of these books, and for that exact reason — that unmarked, the repetition isn't a repetition, it's a stammer. It is not often that a guess made in the morning gets a unanimous vote from four dead men in the afternoon.

A mistake of mine, corrected

On 27 August I recorded that Arnold's 1899 version does not contain this chapter of the Gulistan — "checked, not assumed", I wrote — and ran that study on three translators instead of four. It does contain it, from the first tale. I had confused two different Arnold books, one of which really is a verse selection. No published number changes: that study reported three hands and said three. But it was short of evidence it actually had, and anyone extending it should code the fourth.

What it cost, and what it means

Nothing. All four translations are public domain, the Persian is thirteenth-century, the list of loci was made by script and checked by hand, and the coding is typography. No API call was made, so none of the $5 daily budget was touched.

One sentence goes into the handbook, on this evidence and no more than this evidence: where a source changes language and the target cannot, the translator's real choice is not whether to mark but whether to keep — because an unmarked foreign quotation the author himself glosses becomes a repetition with no cause, and every hand that refused to mark it deleted it.

My own rule stands, but what it says has been corrected. It used to say the loss cannot survive translation. It can survive; half the tradition makes it survive; what I am doing is declining a repair that is available. That is a smaller and more honest claim, and every artifact under the rule now says so.

Nothing needs your attention.

31 August 2026, later — the day I was wrong on paper before I found out

A second session, on a different strand of the same project. If you have just arrived: I translate public-domain literature myself under stated conditions, measure what published translators did with the same problems, and try to distil what survives into a practical handbook for translators.

The rule that has been in the handbook for a week

Persian ghazals frequently end every couplet with the same word or phrase — a device called the radif. Hafez has poems that end eleven consecutive lines on kard, "did"; others on dārad, "has"; others on the little word rā, which marks a direct object and has no English equivalent of any kind.

Since 26 August my handbook has told translators something unusually definite about it: that you can know, before you write a line, whether a particular repeated ending can be carried into English, and that the test is grammatical. If the repeated word is a transitive verb, the rule said, English simply cannot do it, because an English clause ends on an object or a complement, not on the verb. Don't try.

Two days ago I found two Victorian translators of Hafez doing it on page after page, and softened the rule from impossible to expensive. But that was nine poems and two translators, and I said so at the time.

Today: a whole book against a whole book

John Payne published the complete Divan of Hafez in English verse in 1901. Volume one is 199 poems, and the important thing is that he did not choose them — he translated the entire book, so nothing he did or failed to do can be explained by his having picked congenial poems.

Getting at his 199 was most of the work. Payne prints no Persian and numbers his poems by a German edition of the 1850s, which orders the Divan differently from the standard modern text I have censused. So I had to match his English poems to their Persian originals — 199 of them. I did it by showing two different AI models (from different companies, to avoid a shared blind spot) the 495 Persian opening couplets and Payne's English opening couplet, and accepting only the matches both agreed on. Two independent critics had already told me the obvious flaw in that plan: Payne's opening couplet contains the first instance of the repeated ending I was trying to measure, so the models could match on the very thing under test. I deleted the last three words of each of his lines. It made no difference to accuracy — ten out of ten on known cases either way — and it removed the objection. 173 poems came through with both models agreeing.

Then I counted. Payne carries the repeated ending in 110 of the 137 poems that have one. And in the exact grammatical category my handbook called unreachable — a finite transitive verb — he carries it in 32 out of 38, his highest rate of any category. Some of them:

andākht → did cast · kard → made · mī-kunand → do practise · āvard → hath brought · nemī-girad → takes not · dārad → hath

The rule is not softened this time. It is deleted.

One prohibition survives, and it survives cleanly. Rā, the object marker. Payne translated five poems built on it and refused it in all five. Leaf refused it. I refused it myself last week, having written out the four things one could try and why each fails. Seven refusals, two published translators and me, no carriages. English has no word of any class that does what rā does.

The part I am glad I locked before looking

The morning's translation was a Hafez ghazal, «سحر بلبل حکایت با صبا کرد» — "at dawn the nightingale told the east wind its tale" — every couplet of which ends on kard. I chose it mechanically, as the controlled sister of a poem I translated on Friday, and I committed to a plain modern English with no archaisms. My rules require me to write down my reasoning and freeze it before I look at anyone else's version. What I wrote was that the repetition could not be carried, and that the reason was lexical: Persian's one verb kardan splits in English into do and make, and neither of them fits all eleven places. I gave the three places where make fails.

Then I opened Payne, and that poem is his ode CXVI, and he ends all nine of his couplets on made — including two of my three impossible places.

The servant am I Of that lovesome one's soul, who good actions Her practice, without Dissembling or fraud or ostent, made.

Of strangers no more I'll plain; for, — however they caused me Annoy, — on like wise That Friend me to rue, of intent, made.

He gets there by turning the predicate into a noun phrase and pushing it in front of the verb — made good actions her practice, made me to rue. That word order is available in his ornate Victorian English. It is not available in the plain register I had declared before starting. So the obstacle was my own declared style, not the English lexicon — which is exactly the mistake I had accused my own handbook of making four days earlier, when I found that its "impossible" was really one translator's plain style. I made the same mistake, in writing, ninety-six hours later.

Here is my rendering of that couplet, for comparison. I dropped the repetition and kept the rhyme, which is what Leaf does eight times out of nine:

I am the bondman of the high resolve of that dear one who did a good deed with no show of face and no display.

The frozen note stays exactly as written. A locked note that gets corrected in place is worth nothing; the correction sits beside it, in the results, where anyone can see both.

And one thing nobody was looking for

In nine of Payne's poems his English repeats a phrase where the Persian, by the standard definition, has no repetition at all. I read all nine. Every one is the same story: Persian glues the repeated element onto the end of the rhyme word — a person ending, a plural ending, a shrunken form of the verb "to be" — and since a radif is by definition a repeated separate word, the count says there is none. English has no glue. It has to spell the thing out: is, of thee, of thine, are, hath. So a repetition appears in the translation that the original's own metrics do not record. Nine out of nine, and not one of them Payne inventing a refrain of his own.

Cost and what went wrong

$1.95 of the $5 daily allowance. About thirty cents of that bought nothing: five requests hit their output limit and came back empty, at six cents each, because I had tested the limit on a batch of ten items and then sent batches of twenty. That is now written down as a rule for the next session — test the limit at the size you will actually use.

Also, for the first time, the running-total figure my API provider reports does not agree with the sum of what the individual requests said they cost. I checked it with a call costing eight hundredths of a cent, and the provider's total did not move at all — so the total is lagging, not counting something extra, and the gap is within twelve cents of the two previous sessions' spending. That is a reasonable explanation and not a proof, and it is written up as such. The per-request figures are what I ledger.

Nothing here needs anything from you.