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

2026-09-08

This is a long-running study of literary translation: I translate public-domain fiction myself under controlled conditions, have outside AI models judge or measure the results blind, and try to distill what survives into a practical handbook, organized by translation problem.

Closing a five-week question about rhyme, for free

Earlier this month I found something that looked useful for anyone translating rhymed verse: when an English translation of a Persian poem has to rhyme, the very last word of the line is almost always "spent" — it has to serve the rhyme, so it usually can't also carry whatever meaning the Persian original put at the end of its own line. Measured across two published 19th-century English translations of Hafez (54 poems), an ordinary bit of meaning reaches the last word of a free line about 27% of the time, and the last word of a rhyming line only about 2%. I called that "a factor of eleven."

Today's session was set aside to close one loose end from that finding, and it turned up a real correction, at no cost — I didn't need to buy any more AI judgments; I just re-read data I already had, more carefully.

Some of those poems' English translations end their rhyming lines with a short repeated tag — the same word or phrase, like a refrain — echoing a feature of the Persian form called a "radif" (a word or phrase repeated after the rhyme itself, all the way through a poem). In lines like that, the rhyme is not actually the very last word: it comes one or more words earlier, before the repeated tag. My original count only asked "is the meaning at the line's last word?" — which, for those particular lines, is asking about the repeated tag, not about the rhyme.

Redone correctly — asking "is the meaning at the rhyme itself?" rather than "at the line's last word?" — the true rate roughly doubles, from 2.4% to 4.9%. The overall effect is still real and still large: meaning reaches a free line's end about five and a half times as often as it reaches a rhyme, not eleven times as often as I'd reported. Eleven was an overcount caused by measuring the wrong spot in a minority of the poems.

I also have to walk back a smaller claim: I'd said the effect looked the same whether or not a poem used one of these repeated tags, which I took as reassurance that the finding wasn't some quirk of tag-using poems specifically. Corrected, the two groups of poems now look less alike (the tag-using ones show a smaller effect), but there are only eight such poems with enough data to check, so I can't say the difference is real rather than random noise. Correcting an overstatement is a good outcome from a scoring standpoint; failing to be sure whether two groups genuinely differ is an honest limit of a small sample, not a new finding.

What it means for translating: if you're carrying a Persian-style rhyme scheme with a repeated closing tag into English, what's actually constrained is the rhyme word itself — protect that, and don't worry about whatever fills out the tag afterward. The corrected, still-large number (about 5.5 times) is the one that should go into the eventual handbook entry on rhyme, not the original 11.

This closes a set of four sessions Tom asked me to finish before shifting focus back to translating ordinary prose (rather than more Persian verse). The next session moves to that handbook-writing work, and the one after is expected to start choosing a full-length prose work to translate span by span.

Nothing needs your attention.

A second session today: what a translator should — and shouldn't — tell the reader

A second, separate session ran later the same day and wrote a new chapter of the practical handbook this project is building: what a translator prints about the translation itself, rather than in it — a preface line, a footnote, a promise, a title's own wordplay.

Pulling together several earlier findings, four kinds of printed statement turned out to behave quite differently:

• A stated refusal ("I will not attempt to rhyme this") reliably predicts what actually happens on the page. Checked twice, in translations from 1767 and 1850, one held up and one didn't — but a refusal is cheap to check, so it's worth printing.
• A stated promise ("I have kept every pun") is much less trustworthy. In the one case measured — a translator who promised to preserve wordplay in a Persian classic — three independent readers, checking blind, found the promised wordplay nowhere in the fourteen places it should have shown up. (One extra instance turned up later in a spot the original check hadn't thought to look at — so the honest verdict isn't "he lied," just "the readers found nothing where they looked.")
• A comparative claim about oneself ("I repeat rhyme-words less often than my own poet does") turned out to be checkable, and — in the one case tested, a Victorian Hafez translator's own claim — true.
• Telling a reader a plain fact about the source ("this passage rhymes in the original") measurably shifts which of two English versions readers prefer, toward the one that keeps the rhyme. Showing them the original text alongside does not have the same effect — it instead makes readers start checking individual word choices instead of judging the English as English.

This session's own translation — a comic Japanese short story by Miyazawa Kenji (1924) called The Restaurant of Many Orders, about two hunters who wander into a strange mountain restaurant covered in polite instructions, not realizing the "restaurant" is actually going to eat them — put one piece of that guidance to a real test. The story's Japanese title turns on a pun: the word for "order" can mean either a customer's request or an instruction given to someone, and the joke is that all those posted "orders" turn out to be orders given by the restaurant to its guests, not the other way around. The obvious instinct was to add a footnote explaining that pun for an English reader. Drafting the passage first — before deciding — showed that English "order" already carries both meanings on its own, so the twist works without any help:

"All those 'many orders' — it's them placing the orders. On us."

The footnote was written, then deleted, once it was clear the joke survives untranslated.

What it means for translating: printed refusals can mostly be trusted at face value; printed promises should be checked against the actual page before anyone believes them; a fact about the source, stated plainly to the reader, does more work than showing them the original; and before writing a note explaining a joke or device, it's worth drafting the passage first to see whether the target language already carries it without help.

Nothing needs your attention.