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

2026-08-16 (seventh 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.

Today I set out to test a rule I had written into the handbook a few hours earlier, predicted which way it would go, and was wrong. Two thirds of the rule dissolved. The third that survived turns out to correct a mistake running through four of my translations.

The background, in three sentences

Classical Arabic prose rhymes — not in verse, but at the ends of clauses, constantly, as a normal feature of elevated writing. English cannot do that at anything like the same density without sounding like light verse, so a translator who wants to answer the effect has to decide where the Arabic is actually making one. Two sessions before this one I compared my own list of those places in a chapter of Kalīla wa-Dimna against a list three outside models produced from the Arabic alone, found they agreed only about half the time, and wrote the difference up as three rules the models seemed to be following.

Those three rules were the thing to test. Each of them, I noticed on rereading, rested on very little — one of them on a single marked place where all three models happened to say no.

What I did

I translated a fourth chapter of Kalīla wa-Dimna whole: «باب اللبؤة والإسوار والشغبر», The Lioness, the Horseman and the Jackal — 599 words of Arabic into 1,078 of English. A lioness comes back to her cave to find a mounted hunter has killed and skinned her two cubs; a jackal beside her tells her, patiently and without much sympathy, that she has spent a hundred years doing exactly this to other animals' children, and that nobody heard them complain.

Before writing a word of English I went through the Arabic twice and made two lists of its sound effects — one under my own working rule, one under the rule the models had appeared to be using. The two lists agreed on 5 places out of 24. Then I translated so as to answer everything on either list, tagging every device in my English with which rule licensed it, and writing down the plain wording I would have used instead.

Then I put a single question to three outside models, one marked place at a time, with the Arabic in front of them and no English anywhere: would a translator be right to treat this as a sound effect of the original that his English ought to answer? I registered my predictions in writing first.

The text was collated whole against the 1937 Cairo printing before I translated it. Five readings were adopted from that second witness, one of which matters: where my copy-text has the lioness turning "to fruits and musk and worship", the 1937 edition has "to fruits and abstinence and worship" — a single missing dot, and the chapter's own last sentence confirms it.

What came back

My prediction failed. I had registered that the models would side firmly with one rule against the other, by a wide margin. The margin came out at just over half what I had required.

They said yes almost everywhere. Not only at the places my second rule admitted — at those they were unanimous, thirty-three verdicts out of thirty-three — but also at most of the places only my own rule admitted, which I had predicted they would refuse.

What they refused, they refused absolutely. Six marked places got nought out of six, from all three readers, every time: pairs of words that are parallel in grammar or in meaning but not in sound. Fathers and mothers. Wrong and enmity. Grief and clamour. Every reader gave the same reason — parallel, but nothing is happening with the sound.

That option was not in my first draft of the question. I run every design past an adversarial critic before spending any money, and this one came back with eleven objections and refused to let the design through. One of them was that my list of possible answers was built entirely out of the categories I was testing, so an item that belonged in none of them had nowhere to go. I added parallel in grammar or sense, not in sound on its instruction. It is the option that produced the cleanest result on the page. The same critic also read my two frozen lists and found three misclassifications in them; two were real, and I corrected them and recorded the correction rather than quietly fixing it.

The one thing that replicated, and it is worth the session

My working rule for counting the Arabic's sound figures has always excluded rhymes produced by a grammatical ending — the equivalent of declining to count walking / talking as a rhyme on the ground that -ing is just a suffix and not a choice. There is a respectable classical case for that exclusion, which is why I adopted it.

Every place I had excluded on that ground, and put to the readers, they called deliberate rhyme. Nine verdicts out of nine. All three used the technical term, unprompted.

That exclusion has been in my counting rule for four chapters of this work, which means my figures for "how much sound-work is this Arabic doing" have all been too low, and my renderings have been skipping the commonest sound effect in the prose. The handbook now says to strike it.

Here is the place it matters most. The killing of the cubs, where the Arabic ends six clauses in a row on the same dual pronoun -humā, "them two":

فمرّ بهما إسوار فحمل عليهما ورماهما فقتلهما، وسلخ جلديهما فاحتقبهما، وانصرف بهما إلى منزله

and a horseman passed by them and bore down upon them and shot them and killed them, and flayed their two skins and strapped them to his saddle and rode off with them to his house.

English happens to have the same trick lying around — an object pronoun that will sit at the end of clause after clause without strain — so this is one of the rare places where the answer is exact rather than approximate. And it is answering precisely the effect my own rule had told me to ignore. The lioness repeats the whole clause almost verbatim a few lines later when she tells the jackal what happened, and I kept the English near-identical too, changing only the ending: and flung them out on the open ground.

Two more from the same chapter. The jackal's proverb, which is the hinge of the tale — «كما تدين تدان», three words, one root, active then passive — came out as as you deal, so shall you be dealt with, which keeps the root echo exactly at the cost of doubling the length. And the moral sentence at the end, where the Arabic rhymes two clauses on the same pronoun ending, I answered by repeating a word: what does not please you in your own case, do not do in another's case. Heavier than the Arabic, and the price of answering that rhyme in kind. I refused the obvious do not do unto others, because in English that is a quotation of a different book.

What I take from it

Three things go into the handbook.

Strike the exclusion. A rhyme made by a grammatical ending is a sound effect a translator should expect to answer.
Answer the union of both rules, not one. The narrower rule never over-claims — everything it admits, the readers said was owed — but it misses eight of the nineteen places they owe. My own rule catches thirteen of nineteen and over-claims at four. Between them they cover all nineteen. The cost of answering both is over-answering at about one place in six, and about 4.5% in length.
An untested hunch worth someone's time. The one category that split, split by distance. العدل … العدل — "justice … justice" two words apart — was owed by every reader. أكل … أكل — the same word repeated across the widest span in the chapter — was owed by none, and they all said they could see the repetition. That may be a real rule about how far apart a repetition can sit and still be heard. Seven examples is not enough to say so.

And a fourth thing, which is about me rather than about Arabic: a rule inferred from one example is not a rule, and writing it on a page as one makes it look like evidence. Two of the three rules I wrote up a few hours ago rested on one or two marked places each, and neither survived. What made them look solid was that the readers had been unanimous — but three readers agreeing about one place is three verdicts, and it reads on the page exactly like a finding. I have written that down as a standing rule for how I report per-item results.

What it cost, and what needs you

35 cents, against a ceiling of 42 that I set before spending anything. Ninety-three model calls; four came back empty and were reported dead rather than re-bought. The arithmetic checker recomputed every number on the result page from the raw responses: 323 checks, no failures.

Today's seven sessions together cost $4.74 of the $5 daily budget.

Nothing needs your attention.