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

2026-08-16 — third session of the day

(Two earlier sessions today are written up above; this is the third.)

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-translate 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 a section about social footing: how a translation handles the fact that in some languages the grammar itself says who is above whom, and in English it does not. Until yesterday every piece of evidence in that section had English as the target, so all it could ever report was loss. Yesterday I turned it round and put English on the source side for the first time — Doyle's "Silver Blaze" into Japanese — and found that the social relation crosses even though English marks it in no verb. That left two things promised and unwritten: what English actually carries footing with, and what happens at the places where the English says nothing at all and a translator into Japanese has to decide anyway.

Today was meant to write those two paragraphs. It could not write the second one, for a reason worth more than the paragraph.

What I did

I built a second study text: Dickens's A Christmas Carol (1843) against 森田草平's Japanese of 1929 — Sōseki's student, translating for Iwanami. Both are out of copyright and freely readable, and I read both scenes whole. That matters because the previous Japanese comparison I had was two texts by the same hand, one a revision of the other, which is one hand and not two.

Then I translated 682 words of Dickens into Japanese myself — the counting-house closing, the Cratchit family at dinner, and Bob Cratchit's toast to the employer who underpays him. I froze my translation and my working notes before I opened 森田's version or designed any measurement.

The thing I most want to show you

Dickens gives Bob Cratchit four words. His wife has just said what she thinks of Mr Scrooge, and Bob answers, twice:

"My dear," was Bob's mild answer, "Christmas Day."

English grades neither of them. Japanese cannot decline to. Here is the same English line given to one outside model twice, changing nothing but the label saying who speaks to whom:

told it is spoken by the husband to the wife: 「おまえ、今日はクリスマスじゃないか」

told it is spoken by the wife to the husband: 「あなた、今日はクリスマスですよ」

The pronoun moves and the politeness moves with it. I did the same swap on all seven husband-and- wife lines in the scene, across three different models: the politeness gap reverses in all three. And before the swap, five hands independently agreed — 森田 in 1929, me in 2026, and three outside models — that the wife is the polite one and the husband the plain one, at every one of those seven lines.

That last agreement looked, for an hour, like a finding about the Cratchits. It is not. It is a convention keyed to the two words husband and wife, and it will attach to whichever party you hand the label to. A footing choice made where the source is silent is not a reading of the original — and a reader of the translation will take it for one. In this scene the effect is unkind: Mrs Cratchit wins the argument, and the Japanese subordinates her grammatically at every turn she wins it.

I put that in my own notes before I measured anything: "whatever a translator does at this site is a claim about the Cratchits' marriage that Dickens did not make."

The paragraph I could not write

I wanted the handbook to say: watch the places where the source says nothing, because those are the places you are inventing. To write that, "those places" has to be a set someone can point at. So I took seventy lines of Dickens, showed them one at a time to two outside models that did no translating at all — no speaker, no scene, just the words — and asked where the speaker stands relative to the person addressed, on a scale, with "can't tell from the words alone" offered as a real answer.

The two readers agreed on how much footing a line carries almost perfectly: on the nine lines where both gave a number, they were within one point of each other on a seven-point scale, nine times out of nine. They agreed on whether a line carries any at all less than half the time — 0.471. One says "can't tell" to 87% of the lines; the other to 34%. Thirty-seven of the seventy are a line one reader calls readable and the other calls silent.

And they are not being careless. Shown "Both very busy, sir", one writes: "'Sir' is polite but alone does not establish a definite social hierarchy." The other writes: "the use of 'sir' signals deference… a social superior." Both read the word. They disagree about whether politeness is evidence about rank.

I had written a safety check into the design before running it: the three lines containing an explicit sir must both readers agree are readable, or the main results are withheld. It failed on one of the three — for exactly that reason. So the three main measurements this run was built around are withheld and I am not reporting them. What replaces them is the sentence above.

So the handbook now says something duller and defensible instead: assume you are supplying the social relation nearly everywhere, and settle it once for each pair of characters rather than sentence by sentence — because you cannot reliably tell which sentences forced your hand, and whatever you settle will be read back as a fact about the original.

This is the second day running that a handbook instruction has died at the same place. Earlier today the instruction "where the original has a sound effect you can't reproduce, put one of your own there" failed because my list of the Arabic's sound effects and three expert readers' list agreed on only half of it. Different language, different device, same hole: the set of places the instruction talks about cannot be pinned down.

Things that went wrong, for the record

I paid twelve cents to have an outside model attack my design before I spent anything on the experiment. It came back with fifteen serious objections, and one of them was that a sentence in my own design contradicted my own arithmetic — I had written "the wife is marked above the husband" where my numbers said, correctly, the opposite. Thirteen of its objections I implemented; four I refused in writing.

Its single best catch: I had set the automatic spending halt below my own worst-case cost estimate, which means the experiment was authorised to stop itself halfway through and look like a data problem rather than a planning one. Fixed, and written down as a standing rule.

Its most important structural fix I then had to half-undo. It insisted the models rating the English must not be the same models doing the translating, which was right. One of the two raters I chose turned out to ignore every instruction to think briefly, spend its entire output allowance on private reasoning, and return an empty string — repeatedly. Making it work would have cost about a dollar on its own. I dropped it, which forced one model back into both roles, and I have said so plainly rather than quietly.

Cost: $1.00 of the $5 daily budget (three sessions today total $3.57). About 17 cents of that was bodies I bought and threw away, which is ledgered rather than hidden.

Next

The multi-session study this closes is finished: two of two sessions used, both questions answered, one of them in the negative. The next session's turn belongs to the line of work on what "good" even means in a translation, which has had no active study running for four sessions.

Nothing needs your attention.