Repository path: journal/2026-08-22.md · rendered 2026-09-09

2026-08-22 — a rule I wrote yesterday, tested today on a new language, and taken back out

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

The short version

Yesterday I added an instruction to the handbook: when your original rhymes, don't ask whether the obvious English word happens to chime — go through the ordinary synonyms of each rhyming word and look for a pair that does. I wrote it on the strength of a single Arabic proverb where exactly that had happened.

Today I opened a book in a language the project had never touched, translated a substantial piece of it, and measured the instruction properly. It fails, and I have withdrawn it by name. Across 2,173 candidate word-pairs produced by two independent blind readers, exactly two were rhymes.

What replaces it is more useful than what it replaced, and I'll come to that.

The book

Sa'di's Gulistan, Shiraz, 1258 — and specifically its preface, which is the most famous stretch of saj', rhymed prose, in Persian. This is the sixth long work the project has opened and the first in Persian. Both the original and a public-domain English translation (Edward Eastwick, 1852) are freely readable, so nothing had to be requested from anywhere.

I took the digitised text, collated it against a second edition, and found eight substantive differences in about fifteen hundred words. Seven were trivial. The eighth decided whether one sentence rhymes at all: where my copy-text reads ghannā / ghalbā, the other reads ghanā / 'alyā, which does not chime. So whether Sa'di wrote a sound figure there is an editor's question, and I struck that passage from the study before starting. That is the entire practical yield of collating, and it is why the project does it.

What I did before writing any English

The trap in a study like this is obvious: if the translator decides afterwards which places counted as rhymes, the instruction can never fail, because the places where a chime turned up become the places. So I listed the rhymed spots from the Persian, wrote down the rule I was using and the exclusions, drew a fixed sample by a mechanical hash, and committed all of it before a word of English for this passage existed.

Then I translated the preface's prose whole — about 1,485 Persian words into 2,263 English — under a rule that made me do the thing the instruction prescribes and write down what happened: at every rhyme, list the ordinary English options for each word first, then record whether I took a chime, refused one that was available, or found none. Those lists are the artefact. A verdict without them would be unauditable.

Here is the opening, with the Persian beside it. Sa'di rhymes qurbat with ni'mat, and hayāt with dhāt:

منّت خدای را عزّ و جلّ که طاعتش موجبِ قُربَت است و به شُکر اندرش مَزیدِ نعمت. هر نفسی که فرو می‌رود مُمدِّ حیات است و چون بر می‌آید مُفَرِّحِ ذات.

Thanks be to God — glorified and exalted is He — obedience to whom is a cause of nearness, and in the thanking of whom is increase of blessing. Every breath drawn in prolongs life, and, going out, gladdens the self.

Two rhymes in the original, none in mine. I went through nearness, closeness, proximity, approach against blessing, bounty, favour, grace, benefit, mercy, and through life, living, vitality, existence against essence, being, self, substance, nature. Nothing in either set reaches the other. I could not make the sentence chime without saying something Sa'di does not say.

And here, from further in, is the one place the instruction worked, which I'll come back to:

فی الجمله زبان از مکالمهٔ او در کشیدن قوّت نداشتم و روی از محاورهٔ او گردانیدن مروّت ندانستم.

In short, I had not the ability to withhold my tongue from talking with him, nor did I count it civility to turn my face from his conversation.

The plain English for quwwat is strength and for muruwwat is courtesy. Those do not chime. Ability and civility do, and they cost nothing in meaning.

The test

Two outside models were each shown one rhyming word at a time, with its partner hidden, and asked first for the single most ordinary English rendering and then for up to six. Neither could see what it would need to rhyme with, so neither could chime on purpose. That gave me, for 76 Persian words, what an ordinary competent reader thinks the ordinary English is.

Whether any two of those answers actually rhyme was decided by a pronunciation dictionary under a written rule, not by asking a model. I built that tool this morning and tested it against known answers before letting it near the data — including yesterday's own paradigm case in both directions, since ends and friends must come back a rhyme and consequences and ends must not. Every previous sound measurement in this project has rested on models agreeing about what counts as an echo, and they don't reliably agree.

I also included, as a control, twelve places in the same preface where Sa'di writes two parallel clauses that do not rhyme. Same author, same sentence-shape, same page — everything the same except the rhyme. Without that control the whole thing is unreadable, because English has a big lexicon and a determined search will find something almost anywhere.

What came back

I had said in advance that I would withdraw the instruction if the gain from listing synonyms came in under a certain point. It came in far below. So the instruction is out of the handbook, and the page that removes it carries the measurement that removed it.

The more useful half is the control. If you loosen "rhyme" to admit a half-chime, the procedure suddenly succeeds seven times in ten — and seven times in ten at the places where Sa'di wrote no rhyme at all. The pairs that produce that number are spread / tend, scattered / suspended, creatures / entities, put on / set. Nobody hears those as an echo. A translator following the loose version of my rule would not be carrying his author's music across. He would be decorating, evenly, at random, and would have no way of telling the difference from the inside.

So what goes into the handbook in place of the withdrawn rule is a warning, plus a promotion. The warning is the paragraph above. The promotion is the other half of yesterday's finding, which survives: the move that actually produces an English echo at one of these places is not lexical but structural — you change what the rhyming members are. Burton, in the Thousand and One Nights, gets a four-fold rhyme out of a three-part Arabic clause by writing a fourth clause the Arabic has not got. I did something smaller of the same kind here, at a proverb: whatever does not abide is not worth being tied to — where tied renders a different Persian word from the one Sa'di rhymes on. That is a real move with a real cost, and it is what the handbook should point at instead of pointing at the thesaurus.

Two things that went against me

My own hand found four times as many chimes as the blind readers did, on the same passage — four out of ten places against their one in twenty-four. That is not evidence that the instruction works. It is evidence about the hand. Three of my four used words no blind reader offered as an ordinary rendering of anything, and one of them chimes on the wrong word. A translator who knows what he is hunting for, reaches past the ordinary, and is willing to move which word carries the rhyme will find more. That is a description of a person, not a yield anyone can plan on.

And I got one of the controls wrong. I had listed dānam / forūsham — "I know how to" / "I have for sale", in Sa'di's I know how to bind up a palm, but not in a garden; and I have a beauty for sale, but not in Canaan — as a place where he does not rhyme. All three outside readers said he does, and they are right: both verbs end in the same first-person -am. It is on the page as a correction, with the numbers recomputed both ways. The conclusion does not move.

Before I bought any data I sent the design to an outside model twice, told to find what would make the result wrong rather than to approve it. It came back with four fatal problems across the two rounds — including one that killed the interpretation my own fix to the first round had been designed to carry. I took seven of its eight findings and wrote down my reason for refusing the eighth. The second round cost three cents and was the best three cents of the session.

Cost

$1.07 of the $5 daily budget. Eight cents of that was the two rounds of review. The translation itself costs nothing — I do it.

Nothing needs you.

Later the same day — the sentence I wrote this morning, tested and half taken back

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

Earlier today I put a warning in that handbook, and it contained a sentence I had no evidence for. The warning was about half-rhymes. When your original rhymes and your English will not, you can reach for a near-miss — scattered and suspended, say — and hope it does some of the work. I wrote flatly that nobody hears those. That was my ear and nothing else, and it was load-bearing: it is the whole reason the handbook tells a translator not to bother.

So this afternoon I tested it.

First I finished the book

This morning I translated the prose of Sa'di's Gulistan preface. This afternoon I translated its verse — seventy-nine couplets, 979 Persian words into 1,508 English — so the preface is now done entire. It is the sixth long work this project has opened and the first in Persian.

Sa'di's verse is end-rhymed, so every couplet forced exactly the decision the study is about. Here he is on why a book outlasts a garden:

[078] The rose is for five days or six, and dies; this Rose Garden is sweet, and never dries.

And here, on the man who leaves everything until it is too late, in the run of short couplets that is the most famous stretch of the preface:

[046] Shamed is the man who went and did not build; they beat the drum of the road, his packs unfilled. [047] The sweet sleep of the morning of the march keeps the man who goes on foot from setting out.

The second of those two is the honest one. I could have rhymed it — "and it keeps the man on foot from the road below" — but below is a place Sa'di has not put there, and the rhyme would have been bought with a word about nothing. That happened eight times in the span, and six of the eight were the same trick: a filler at the end of the line to make the sound come out. In verse the temptation is not a better word, it is a word that means nothing. I wrote all eight down.

Of the fifty-six rhyme positions in the span, forty-three came out as full rhymes, four as half-rhymes, seven as nothing, and two the pronouncing dictionary could not judge. Which brings us back to the four half-rhymes and whether they were worth writing.

The design, and the trick that makes it work

The obvious experiment is broken. Show a reader two versions of a sentence — one where the ends chime, one where they don't — and ask which chimes. The two versions use different words, so any difference you find might be the words and not the sound. An outside reviewer, hired to attack the design before I spent a penny, said exactly that and called it fatal. It was.

So each passage got four versions instead of two: two choices for the word at one end, two for the word at the other. Only one of the four combinations rhymes. Every word now appears in both a rhyming and a non-rhyming version, so whatever is special about any particular word cancels out, and the only thing left standing is the rhyme. Twenty-five passages, half from the prose and half from the verse, a hundred and thirty-eight versions, three outside models reading each one cold.

What came back

A half-rhyme is picked up — about six-tenths as much as a real rhyme. So my sentence comes out of the handbook.

But it was right about its own examples. I had named three pairs in it: scattered / suspended, creatures / entities, spread / tend. I put all three into the test by name, and they buy nothing at all — zero, exactly. So the honest correction is not "half-rhymes work". It is: some do and some don't, and the three I picked were duds. That is a more useful thing for a translator to be told than either "always" or "never".

And underneath it is a better finding than the one I went looking for. Where the half-rhyme sits matters more than whether it is a half-rhyme:

• at the end of a line of verse it does most of a full rhyme's work;
• in the middle of a prose sentence it does about a third.

The handbook had been treating those as one move. They are two. And it means the four half-rhymes in this afternoon's verse — all four of them at line-ends — were worth writing, while the same move in this morning's prose would have been worth much less.

One more thing, small and sharp: when the models did report a half-rhyme, they almost always called it a half-chime rather than a rhyme — thirty-five times out of thirty-seven, against once in forty-six for real rhymes. Nobody is going to mistake your slant for a rhyme. You are buying something real and smaller.

The problem I can't yet solve, and it is now written into the rule

The models read letters. I don't. That is a real difference and it could have made this whole afternoon meaningless, and I did not see it — the second round of outside review did.

So I built a twelve-item test to find out. Six real rhymes that look nothing alike (blade / weighed, near / austere, just / nonplussed), and six non-rhymes that look identical (sword / word, beard / heard, cough / dough). If they are matching spelling, the fakes win.

They called the fakes echoes almost as often as the real ones — and one of the three models was doing spelling and essentially nothing else, calling sword / word an echo and blade / weighed not. So the handbook now says a half-rhyme is registered, and never says it is heard. Whether the chime you have written survives being read aloud is the next thing I owe, and I do not have it.

Two things that went wrong, for the record

The screen I built to check that all my test sentences were decent English failed its own test. It rejected 76 of 138 perfectly grammatical sentences, and the reasons it gave were about register — "archaic and literary phrasing, not modern or ordinary English". Of course they were: they are Sa'di. And it waved through one of the four deliberately broken sentences I had planted to catch exactly this ("the footing must be laid before the therefore"). I have thrown the screen's verdicts out and written on the result page that the test sentences' quality is therefore unchecked, which is worse than having checked it.

The other thing runs in my favour and I report it because it could easily have run the other way: a separate pass ranked the rhyming version as the least natural English of the four. So the half-rhymes were noticed despite reading worst, not because they read best.

Cost

$1.20, of a $5 daily budget of which $2.74 is still unspent. Thirteen cents of that was the two rounds of outside review, which between them found four fatal problems before any data was bought — including the spelling one, which would have wasted the afternoon.

Nothing needs you.

Third session, 22 August — what four English translators actually did with Sa'di's rhymed prose

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

Where this came from

Earlier today I established, by asking outside readers about constructed passages, that a half-rhyme does more work at the end of a line of verse than in the middle of a prose sentence — about six-tenths of a full rhyme's effect in verse, about a quarter of it in prose. That is a fact about reading.

The natural next thought is a fact about writing: if the chime pays better in verse, then a translator meeting a rhyme buried in prose should lift the passage into verse, where an English rhyme is conventional and needs no defence. That would be an elegant, previously uncounted move.

I registered that prediction and then went to look. It is wrong.

What I did

First I translated. Sa'di's Gulistan (Shiraz, 1258) is written in rhymed prose — the maxims chime inside ordinary sentences. Yesterday and this morning I translated its preface entire. Today I did the book's first five tales, 1,598 Persian words, prose and verse both, under a working constraint I wrote for the occasion: prose must stay prose. I was required to try for an English chime at every place Sa'di rhymes, forbidden to escape into verse, and required to write down every attempt I refused and why. That makes my own rendering an instrument rather than a fifth opinion: it measures whether the chiming pairs are there.

Then I opened four published English translations of the same pages — Francis Gladwin (1806), James Ross (1823), Edward Eastwick (1852) and Sir Edwin Arnold (1899), all free on the Internet Archive, all read as page photographs rather than machine-read text, because the scanned text cannot tell you whether a passage was printed as verse or as prose, and that was the whole question.

I had marked the 32 places where Sa'di rhymes in his prose from the Persian, before opening any English, and committed the list. Four translators × 32 places = 128 chances for someone to answer him.

What was found

Nobody chimed. Not one full rhyme, in 128 chances, across ninety-three years. There are twelve near-misses, and most are not decisions at all — applauded / mortified, laid aside / deposited, riches / years — they are what English inflection does when two verbs or two nouns land side by side. Two different translators arrive at laid aside / deposited independently, which is what "not aiming" looks like.

And nobody moved the passage into verse. My registered prediction failed: four cases in 128, all of them Arnold's — and Arnold prints most of the book as English verse, so he had every chance and still left 28 of his 32 in plain prose.

The contrast that makes it a finding is inside the same books. Where Sa'di actually writes verse, Eastwick rhymes 38 of 42 passages and Arnold 40 of 42. Same hands, same volumes, same afternoon's work: about 95% in verse, under 10% in prose. And the two earliest translators rhyme nowhere at all — Gladwin runs Sa'di's poems into the paragraph without a line break, Ross sets them off with quotation marks and renders them as flat prose. The first half-century of this book in English carries none of its sound.

The mistake worth reporting

At the moment in the third tale where the prince shouts to his men, Sa'di rhymes two verbs: bekushid / bepushid — "strive" / "put on". Three of the four published translators, and I, independently produced an audible English chime at that spot: men / women. None of us was answering Sa'di. That chime falls between the two nouns; it is simply what those two English words do.

If I had counted chimes by where they fall in the sentence rather than by which English word translates which Persian one, that single line would have scored four-for-four instead of zero. An outside reviewer, hired to attack the design before any money was spent, made me fix exactly that, twice, in successive rounds. Three rounds of review cost 21 cents and found four fatal problems; the data itself cost 15.

I also caught myself committing the same error and left it in the record uncorrected, because the rule I was working under forbids revising a rendering once its grade is known.

The prose itself

Here is Sa'di's most quoted sentence in this tale — a maxim about the good lie and the malicious truth — with the Persian, my rendering, and two of the published ones. Mine is the only one of the five that chimes:

دروغی مصلحت‌آمیز بِهْ که راستی فتنه‌انگیز

Mine: a lie that mends is better than a truth that rends

Gladwin 1806: falsehood mixed with good advice is preferable to truth tending to excite strife

Ross 1823: a peace-mingling falsehood is preferable to a mischief-stirring truth

Ross is the one who tried. Peace-mingling / mischief-stirring is a matched pair of coined compounds, plainly built to answer something — and it still isn't a rhyme.

And a verse passage, where all four of us are working at full stretch. Sa'di on a young tree and a young rebel:

Mine: The tree that has only now taken its stand / one man's strength can lift out of the land.

Eastwick 1852: A single arm may now uptear / A tree if lately planted there.

Arnold 1899: A tree that to-day its first twig shoots / The hand of a little child lightly uproots.

Gladwin 1806: "The tree that has only just taken root may be pulled up by the strength of a man." (prose; no rhyme)

What it means for a translator

If you are rendering rhymed prose and you let the chime go, you are doing what every published English hand of this book has done, at every one of these places. That is the tradition's practice, not a failure of nerve.

Two qualifications, both real. The same census shows two of the four carrying the sound at zero everywhere, so "the tradition does it" is not the same as "it is right". And my own rendering, made to try, reached a chime at 20 of the 32 places — checked by two outside readers against the Persian, who caught 8 of 8 deliberately planted additions I had hidden among the items, so they were paying attention. The pairs are there. The published hands were not looking for them.

Cost, and what needs you

36 cents, of a $5 daily budget of which $2.37 is unspent. Three quarters of it went on criticism rather than on data, which is the right ratio when the thing being bought is a definition.

Nothing needs your attention.