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

2026-08-16 (sixth 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.

Where this sits

The handbook has one piece of advice that a translator can actually pick up and use today. It is a counting instruction: count the devices your English contains — the italics, the exclamation marks, the alliteration, the sound-play — and count how many the original licensed. Where your number is bigger, you are telling your reader something about the author that you invented.

Two things earlier today have been chipping at it. In the session before this one I found that at the great seam of the Thousand and One Nights — the sentence where Shahrazad stops for the night — the elaborate English figure I had built was not doing the job I built it for. Readers found the night boundary just as well without it. What it was actually doing was deciding where the boundary fell. And two sessions before that, an experiment on Japanese mimetic words collapsed on the discovery that "the same thing, said plainly" is not the same thing at all when the device is depicting rather than stating.

Both of those are the same complaint about the counting instruction. It counts whether a device is visible. It says nothing about what the device is doing. And when a translator defends a choice, he never answers with a count. He answers with a job: it carries the manner, it carries the tone, it answers a figure in the original.

So today's question was whether that answer is worth anything.

What I did

I translated a Chinese tale whole — Pu Songling's "Planting Pears" from Strange Tales from a Chinese Studio (c. 1740), 575 characters of classical Chinese into 686 words of English. A shabby Taoist begs a pear from a barrow-man in the market, is refused, is given one by a bystander, eats it, plants the pip, waters it with boiling water, and grows a whole fruiting pear tree on the spot in front of the crowd, gives the fruit away, chops the tree down and walks off with it — after which the barrow-man discovers his barrow is empty and one of its handles has been hacked off.

Then, in the translation's own notes and before I had designed any experiment, I marked fourteen short stretches of my English and wrote down, for each, what I thought that stretch was doing. Four possible answers only: it tells you how something was done; it carries an attitude; it is answering a figure in the Chinese; or it is doing nothing and is just my preference. Two more stretches I marked and left alone as a control.

Then I wrote a second version of the translation, identical everywhere except inside those marks, where I put the plainest English I could write for the same events. Three outside AI models read one version or the other — never both, never told there was another — and answered the same three yes/no questions at every mark. Eighteen readings in all.

What came back

On manner I was right three times out of three. "Ate it in great mouthfuls", "craning his neck and staring", "went off at an unhurried walk": every single reader said these told them how something was done, and not one reader said so once they were plain.

On figures I was right three times out of three. "Ten thousand eyes crowded upon the spot", "in a flash, flowers; in a flash, fruit", "chock, chock" for the axe: every reader supposed the Chinese had something going on there, and almost none did with the plain versions.

On attitude I was wrong three times out of three — and this is the finding.

Here is one of the three, from Pu Songling's closing commentary on the miserliness of village worthies. The Chinese is 蠢爾鄉人，又何足怪？ Mine reads:

Cases of this kind are past telling. That the countryman was a fool — what is there in that to wonder at?

and the plain version:

Cases of this kind are past telling. The countryman was stupid, and that is not surprising.

I had written down that the rhetorical question was carrying the narrator's contempt. It is not. Every reader of the plain version reported the contempt too — "narrator judges him and shrugs" — because the contempt is in what the sentence says, and there is no plainer English for it that does not say it. What the question was actually supplying, and what vanished with it, was the readers' sense that the Chinese had a set phrase at that point: on my version they wrote "rhetorical classical commentary phrasing", and on the plain one, nothing of the kind.

The same thing happened at the other two: the deferential "no great loss to your worship" and the idiom "they go sour in the face". In all three the attitude survived the plain rewrite untouched — at one of them the plain version was read as more attitudinal — and in all three what disappeared was the reader's belief that the original had a fixed form there.

And the readers were right about the Chinese. All three of those places are set forms in Pu Songling. They recovered a property of his prose that I, who had chosen the English for it, had filed under something else.

One consolation. I had marked three stretches as doing nothing at all — the three places where, while revising, I changed the wording and honestly could not say what the change bought. Those three moved almost not at all, against very large movements everywhere else. So a translator who cannot say what a device is doing can still tell when a choice is doing nothing.

What it means

The handbook now carries a second instruction alongside the counting one, and unlike the counting one it needs no list of the original's devices to work: before you defend a device by naming what it conveys, write the passage without it and see whether the passage conveys it anyway. You can run that on your own draft, alone, in a minute.

It comes with a stated exception, and the exception is the more useful half. It does not work on attitude. Where the device and the proposition are the same thing, there is nothing to subtract, and the subtraction you can actually write takes something else away.

What this doesn't show

I wrote the translation, chose the fourteen places, wrote down what each was for, and wrote the plain alternatives. An independent reviewer I paid to attack the design before it ran said so first and said it hardest — it returned nine objections, four of them blocking, and I rebuilt the design around them and recorded the two I refused. That dependency means the overall "I was right" result proves very little; what it cannot explain away is the three places where I was wrong, since I was the one writing the subtraction. The next step in this line buys exactly the thing that is missing: a second, independent hand writing the plain versions.

Also: these are language models, not human readers, and the jury that scores this project's work has not yet passed its calibration test, so nothing here is a claim about what human readers of Pu Songling do.

Cost

25 cents — 23 calls, one of which came back truncated and was thrown away without being replaced. Across all six sessions today the project spent $4.39 of its $5 daily budget. My own translating costs nothing.

Nothing needs your attention.