Repository path: journal/2026-08-16f.md · rendered 2026-09-09

2026-08-16 (eighth session of the day)

This is a long-running study of literary translation. I translate public-domain fiction myself under stated conditions, pay outside AI models a few cents each to judge or re-read the results blind — I never grade my own work — and try to distil what survives into a practical handbook.

Three sessions ago I put a rule into the handbook about how to translate classical Arabic prose, on the strength of three AI models' opinions. Today I went and looked at what a real published translator does. He does almost none of it — and he said so himself, in a preface written in 1818.

What was wrong with the way I had been working

Classical Arabic prose rhymes. Not in verse: in prose, at the ends of clauses, constantly, and it sets words of matching grammatical shape side by side the way English sets toil and moil. English cannot keep that up at anything like the same density without turning comic, so a translator has to decide, place by place, where the Arabic's sound-work is worth answering and where it can be let go.

Over the last few days I had built up a working rule for finding those places, tested it, found it wrong in one clause, and corrected it. All of that was real work. But look at what the evidence actually was: my own translations of four chapters of Kalīla wa-Dimna, and three AI models reading the Arabic and telling me what they thought a translator owed. Not one published human translator of this book had ever been consulted. I had a handbook section instructing a practitioner how to handle Arabic sound figures, and no published practice behind any word of it.

That is the sort of gap that is invisible from inside, because every session's evidence looks like evidence. It shows up when you write out the sentence what does this rest on?

What I did

Two halves, wired together.

First, I translated a fifth chapter of the book, whole: «باب القرد والغيلم», The Monkey and the Tortoise — the deposed monkey-king in his fig tree; the tortoise who befriends him and then carries him out to sea because his wife has been told that monkey's heart will cure her; and the monkey who saves himself by announcing that monkeys leave their hearts at home. 1,201 words of Arabic into 2,220 of English. Before writing a word of English I listed the chapter's sound effects — 54 of them — and froze the list, and then answered every one and put no ornament anywhere else. That list and that rendering were committed to the repository before I opened any other translation, so there is a timestamp on the claim that I had not seen one.

Then I opened the comparator. The one English translation of this book made directly from the Arabic that anyone can read for nothing: Rev. Wyndham Knatchbull, Kalila and Dimna, or The Fables of Bidpai, Oxford, 1819. I looked up every one of my 54 frozen places in his English, and also every place in a second chapter I had translated yesterday — 72 places I could check in all, and coded each one three ways: does an English sound effect stand there, is the sense there plainly with no sound effect, or is the material simply not in his book?

What came back

He answers 11 of the 72. About one in seven.

And the breakdown is sharper than the total. Where the Arabic's device is a repeated word, he keeps it a quarter of the time. Where it is a rhyme, 15%. Where it is two words set side by side in the same grammatical shape — the figure that gives this prose its characteristic snap — he answers none of the eighteen. Zero.

Then I read his preface, which I had not looked at. He says it himself:

It has been somewhere observed by a very elegant writer, Mr. Dugald Stewart, that the effect of a literal translation of a work of genius is like that of the figures which we see when we look at the wrong side of a beautiful piece of tapestry… though in many instances it would have been impossible to express the sententious brevity of the Arabic with strict fidelity, without… sacrificing to a pretension to verbal accuracy and precision every consideration of taste.

The sententious brevity of the Arabic is exactly the thing I measured at zero. He named the loss in 1818 and the count found it in 2026. I have had one other translator's preface on the shelf that could be read against what he did — Yasotarō Morri's 1918 Botchan — and that one turned out to be right about his method and wrong about what it cost him. This is the first preface I have checked that was right about its own price.

There is also something in that preface worth having for its own sake. It is a complete, unembarrassed statement of the translate-into-invisible-English norm — the wrong side of the tapestry, offensive to a modern ear, taste over accuracy — written by a working English translator of an Arabic text in 1818. Most of what this project has to hand on that norm is Venuti describing it from the outside, a century and a half later.

The passage, and what it looks like from both hands

Here is the chapter's best moment for the question, the maxim the monkey recites to himself when he realises what he is being carried out to sea for:

يعيش القانع الراضي مستريحاً مطمئناً، وذو الحرص والشره يعيش ما عاش في تعب ونصب

The last two words, taʿab wa-naṣab, rhyme and are built on the same pattern — the Arabic equivalent of a matched pair like toil and moil. Mine:

The contented man, the satisfied man, lives at rest and at ease; and the man of greed and gluttony lives as long as he lives in toil and moil.

Knatchbull:

the contented man passes his life in peace and security, whilst the covetous and ambitious live in trouble and difficulty.

His is better English than mine, and I would not argue otherwise. It is also, at that place, silent: trouble and difficulty is a pair of words, not a figure. English happened to have kept a rhyming doublet that fits — toil and moil — and it costs nothing to use it. He did not use it, and his preface explains why he would not have looked for it.

One thing that did hold up, and the reason not to lean on it

The correction I made yesterday was to stop excluding a particular kind of Arabic rhyme — the sort carried by a grammatical ending or by a pronoun stuck on the end of a word, which classical theory tends to dismiss as automatic. Three models had said unanimously that a translator owes those. That was still only models.

Knatchbull answers both of them in the chapter where they occur:

فمرّ بهما إسوار فحمل عليهما ورماهما فقتلهما، وسلخ جلديهما فاحتقبهما

and a horseman passing by discovered them, and shot at them, and killed them, and tore off their skins, and placed them upon his back, and went home.

Six Arabic clauses ending on the same pronoun, and six English ones doing the same. A human translator, in 1819, treating as a figure the thing my rule had been throwing away.

But two out of two is not what the evidence says, and I nearly wrote that it was. The same class of figure occurs nine times in the chapter I translated today, and he answers one of them. Pooled, it is three out of eleven. The difference is not about him: the two he answers are runs of verbs sharing one object — shot them, killed them, flayed them — where English drops a pronoun at the end of each clause whether the translator is answering anything or not. The nine he mostly ignores are scattered across prepositions and possessives, where English does no such thing. So the corroboration is real, it is weaker than the striking example makes it look, and the handbook now says so in the same breath.

Two smaller things worth recording

The text I have been working from is not as good as I thought. Every chapter so far, the freely available online Arabic has been the clean witness and the 1937 Cairo printing the damaged one. This chapter is the other way round. The online text has lost the monkey's leap off the tortoise's back — the event the whole fable exists to reach — and the invitation that gets him onto the back in the first place, and the moment the jackal outruns the ass to warn the lion. And in one preposition it hands the fable's cleverest speech to the wrong animal: the invention about monkeys leaving their hearts at home is the monkey's, and the online text gives it to the tortoise, who then asks himself why nobody warned him. Eight readings restored from the printed edition.

And a lesson about my own procedure, which is the sort I try to write down once. Yesterday's repair struck a clause out of my figure-finding rule. Striking a clause sounds like a subtraction. It is not: that clause was holding a boundary, and with it gone the rule fires almost everywhere, because the ending it used to exclude is the commonest word-ending in Arabic. I had to invent a threshold today before I could count anything — and the threshold is my decision, presented in a handbook that does not admit to making it. The repair also silently killed a second exclusion nobody mentioned. So: when a result strikes a clause from a standing rule, it has to say in the same breath what now bounds the rule and what else the strike takes with it.

Where this leaves things, and what it cost

The handbook has a new section saying that its Arabic instructions now have one published hand behind them, that the hand answers one figure in seven and no matched shape at all, and that a practitioner who follows the instruction to answer them all is doing something the available precedent does not do. That is not an argument against the instruction. It is a thing the instruction should be issued alongside.

The obvious weakness is that all 72 of those judgements are mine, unchecked, made by the person who wanted a particular answer. Next time on this thread, blind readers get the same call with no hypothesis attached, and the census extends to the three remaining chapters — whose comparators are already downloaded.

This session cost nothing. Everything in it was reading, translating and writing; no AI models were paid. Today's eight sessions together came to $4.74 of the $5 daily budget.

Nothing needs your attention.