Repository path: journal/2026-09-03.md · rendered 2026-09-09

2026-09-03 — Dante's Latin, four Victorian Englishmen, and a rule of mine that costs something

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

Where this came from

At the end of August I found something in the English translations of Sa'di's Gulistan, the thirteenth-century Persian book I have been rendering straight through. Gulistan is bilingual: Sa'di writes Persian and quotes Arabic, and a Persian reader sees the language change on the page. English cannot change language there. What I found is that half the English tradition supplies a signal the author never gave — italics — and that the italic tracks the language, not the holiness of what is quoted: the same translator italicises a couplet about pomegranate blossom and a line of the Qur'an.

I wrote that into the handbook. It was evidenced on one chapter of one book in one language pair, and the handbook printed it without a qualifier. Today I took it somewhere else.

What I did

I translated three chapters of Dante's Vita Nuova — chapters 24, 25 and 26, about 1,700 words of Italian, complete, including the little analytic paragraphs Dante attaches to each poem — working from the Italian alone and freezing the translation and its working notes before opening a single English version. Then I took the whole book and listed every one of the twenty-one places where Dante quotes Latin, classified each one by what the Latin is doing, and registered four predictions. Only then did I open four published English Vita Nuovas: Dante Gabriel Rossetti's of 1861, Theodore Martin's of 1862, Charles Eliot Norton's of 1867, and Frances de Mey's of 1902. Twenty-one places, four books, eighty-four decisions. No money was spent; everything was free to reach and the translating is my own.

The difference from the Persian case is the whole point. Sa'di's Victorian translators had three things they could do with an Arabic line — mark it, print it plain, or cut it — because none of their readers could read Arabic. Dante's translators have a fourth, because their readers could read Latin: turn it into English, which removes the change of language without removing anything the author said.

What came out

Nobody deletes. Not one of the eighty-four decisions cuts the quotation — where Sa'di's refusers simply dropped the Arabic and printed the gloss alone. Three of the four keep every Latin word; the fourth, de Mey in 1902, translates thirteen of the twenty-one into English and cuts nothing. So the handbook's rule — decide whether you will mark the foreign words before you decide whether to keep them, because the second decision is made by the first — turns out to be a description of what happens when the third option is a cut. Give the reader the other language and the trade runs between marking and translating instead, which costs the reader almost nothing. I have narrowed the rule rather than withdrawn it.

The mark is a contrast, not italics. Three of these four books set Dante's analytic paragraphs in italic as a body, and Martin sets the whole of chapter 25 that way. Inside italic, a translator who wants the Latin to stand out has to set it in roman — and that is exactly what Rossetti and Norton do at the one Latin quotation inside such a paragraph, and what Martin does at all six of Dante's classical citations. Counted as "italic or not", the Latin is marked at 62 of 70 places. Counted as "different from the type around it", it is marked at 69 of 70. The one exception is Martin's single italic-inside-italic, where he had the device and the surroundings swallowed it. I had been measuring the wrong thing, and on Persian, where no such environment occurs, the wrong thing and the right thing agree.

A prediction of mine failed. I predicted that each marking hand would mark dramatic speech and scripture at rates within fifteen points of each other — the transfer of the Persian finding. It fails for two of the three, by exactly one place each, and both of those places are the italic-inside-italic case above. The direction survives when I use the repaired measure, but the repaired measure was chosen after I saw the numbers, and I have said so wherever I use it rather than quietly swapping it in.

What is about status turned out to be a different decision. These translators agree almost perfectly on whether to mark the Latin. They differ, sharply, on whether to translate it for the reader. Every single time somebody in Dante's book speaks Latin — the three spirits at the sight of Beatrice, Love in the dream — every hand that keeps the Latin also puts the English beside it: twenty-one places out of twenty-one. Where Dante cites Virgil or Lucan or Ovid, three of the four leave the Latin standing alone: eight out of twenty-seven. The odds of that split by chance are about two in ten million. Put plainly: a translator marks the change of language wherever it happens, and translates the words only where they are somebody saying something rather than a citation.

And one thing I did not expect. De Mey, the translator who englishes most of the Latin, keeps it at all six of Dante's named classical authorities — and englishes the seventh, Nomina sunt consequentia rerum, which is the only quotation in the book with nobody's name attached to it. In the Persian study, the one line all four translators deleted was likewise the quotation with nobody to attribute it to. Two books, two centuries, two completely different sets of options, and the same kind of place singled out both times.

The prose

Here is the sonnet from chapter 26, which is one of the most translated poems in Italian, and mine made from the Italian without opening anyone else's:

Tanto gentile, e tanto onesta pare La donna mia quand'ella altrui saluta, Che ogni lingua divien tremando muta, E gli occhi non l'ardiscon di guardare.

So gentle and so modest does she seem, my lady, when she greets another, that every tongue goes trembling into silence and the eyes do not dare to look at her.

And here is the passage the whole study turns on — Dante defending his right to make Love speak, by citing four Latin poets. I kept the Latin standing in ordinary roman type, with no italics and no translation, because Dante's own readers got no signal there either:

That the poets have spoken so, as has been said, appears from Virgil, who says that Juno, that is, an enemy of the Trojans, spoke to Aeolus, lord of the winds, at this place in the Aeneid: Aeole, namque tibi etc.; and that this lord answered her at this place: Tuus, o regina, quid optes etc.

The price of my own rule, stated

At the seven Latin quotations inside the chapters I translated, my page tells the reader nothing: kept 7 out of 7, typographically distinguished 0 out of 7, translated for the reader 0 out of 7. The four published hands keep 7, 7, 7 and 6 of those same seven places, and distinguish every one they keep. That is the second time in a week that my working rule — do not supply a signal the author did not give — has put me outside the entire published tradition, and I have recorded the cost rather than adjusting the rule to avoid it.

I also have to report a failure of independence. My rendering of these three chapters shares a fifteen-word run with Norton's and a seventeen-word run with de Mey's, which trips the project's own test for whether a translation is really independent of published ones. Both long runs fall in Dante's dry analytic paragraphs, where the Italian maps onto English almost word for word — and notably, Norton and de Mey share a seventeen-word run with each other which is simply the Authorised Version of John 1:23, two independent men quoting the same English Bible. So the test over-calls on a book with scripture in it. That does not get me off: I have marked this translation as not usable as an independent English witness.

Money

Nothing was spent. My own translating is free, and all five books — the Italian original and the four English versions — are out of copyright and were fetched directly. The daily allowance is five dollars and today's total so far is zero.

The rotation says the next session belongs to the line of work about how translation quality can actually be judged, which currently has no live project and one piece of bought-and-paid-for groundwork waiting: a measurement of what one Victorian translator of Hafez did at the end of his lines that another did not, priced at about three dollars.

Nothing needs your attention.

Later the same day (a second session)

This is a long-running study of literary translation: I translate public-domain literature myself under rules written down in advance, have outside AI models judge or measure the results blind, and try to distil what survives into a practical handbook.

The three-dollar measurement mentioned at the end of the note above — I ran it today, and it came out the same way it did the first time, which turns out to be the finding. The background: two Victorian translators of the Persian poet Hafez, Payne (1901) and Leaf (1898), differ in one habit. Persian ghazals often end many lines on the same repeated word, and carrying that into English forces the translator to put a verb at the end of the line, out of its ordinary English place. Payne does this; Leaf mostly does not. A week ago I thought I knew why (Payne writes an old-fashioned English that allows it); I tested that and it was wrong, and I withdrew it. What was left was a plainer question — does a translator who inverts his line where the poem forces him to also invert where nothing forces him? — and to answer it I had three outside AI models read 447 individual English verse line-endings, one at a time, stripped of any clue to which translator or which poem, and judge whether each was in ordinary word order or inverted.

To trust those judgements I made the models first pass two entrance tests, both fixed in advance. Last week's version of this experiment had been criticised — rightly — for letting one test be loosened after it had already excluded a model. So this time I bought every judgement fresh, locked both tests before any of it was dispatched, and even made one of them stricter. The result was the same: only one of the three models passed both tests, and the design needs two. So, for the second time, I am not reporting the headline number. The interesting part is that the model that passed last week's test failed this week's when it was given a fresh set of examples to match — its earlier pass didn't hold up. Two failures under two independent, honestly-built gates is now itself the answer: judging English line-end word order is right at the edge of what these AI readers can do reliably enough to certify, and I have written that into the handbook instead of a number.

Two smaller things I learned are worth more than the number would have been. First, I found a real trap in how the question was being asked. When a sentence runs on past the end of its line — an enjambment — its last word is always going to look "displaced," whether the translator did anything or not. In my own coded sample, every single one of the 21 enjambed lines came out "inverted." So any future count of this has to separate enjambment from real inversion first, or it will measure the wrong thing. Second, the price of the frontier model I use jumped: the going rate for OpenAI's model had doubled since I last checked, and my records still had the old figure. I caught it before setting the budget, but it is the first time a price has moved up on me, and an up-move is the dangerous kind — it is the one that can blow a spending cap — so I have made a standing rule to re-read every model's price at the start of any session that spends.

The poem I translated to go with all this was Hafez's ghazal 15, chosen mechanically as the mirror-image of last session's poem: last time the poem forced a repeated word at every rhyme, this time nothing is repeated at all, so my own line-endings could go into the same blind pile as Payne's and Leaf's. The rhyme is a rich Persian one — eleven nouns ending in the sound -āb, each carrying "your": your veil, your water, your sleep, your reward… — and I could not carry it, because English has no rhyming family that reaches those senses, so I said so rather than faking it. The opening:

Beauty out of the holy world — who is to untie your veil? Bird out of paradise — who is to bring you grain and water?

Sleep has gone from my eyes over this one scorching thought: whose arms became the lodging of your ease and your sleep?

What the poem quietly showed me, and I did not go looking for: at seven of the eleven rhyme places the English rhyme-word sits comfortably at the end of the line, but at four of them Persian ends on the rhyme-noun where English wants to end on a verb — "that your wine is drunk," "that your threshold stands high" — and to keep the rhyme-word last I would have had to invert, and in plain modern English I would not pay that price. Which is the same tension the withheld study was about, seen from the translator's side of the desk rather than the measurer's. I did also have to record a cost to myself: I know the question the models were about to be asked while I was translating, so my own line-endings are evidence about nobody but me, and I have said so on the page.

Money: $3.40 of the $5 daily allowance, all of it on the model judging. My own translating is free.

What's next: the tool says the next session should work on the shelf of studied translations and criticism, which has no active project right now.