Repository path: wiki/findings/results/RS-20260831-arabic-in-persian.md · rendered 2026-09-09

When the source is bilingual and the target cannot be: four English hands on Sa'di's Arabic

The question. «گلستان» is written in Persian and quotes Arabic — Qur'an, hadith, Arabic verse, Arabic proverb — and a Persian reader sees the language change on the page. English cannot change language there, because the whole book has already become English. What do published translators actually do about it?

Why it was asked today. The translation limb of the same session rendered حکایات ۲۱–۳۱ whole (T-gulistan-bab2-R05-v1, span C) and met the problem four times, once at a Qur'anic half-line that rhymes with the Persian half-line before it (b236, زِنهار / النّار), where the code-switch is what makes the rhyme. The arm's standing rule V5 says the English marks the Arabic only where Sa'di marks it, and that where he does not, the loss is declared rather than repaired. This study asks whether that is the tradition's rule or only this arm's.

Design frozen before any English was opened: loci-frozen.md, committed at 0326b5d6, after the translation was committed at 651d64cc. Coded cells with the English quoted verbatim: coding.md. Arithmetic: analyse.py → analysis.json. Cost $0 — no API call was made.

1. What was counted

21 Arabic quotations enumerated from the Persian across the whole chapter by two independent screens and then read and classified by hand; 17 unframed, 4 framed, one of the four also glossed by Sa'di into Persian in the same sentence. Pious formulae, naturalised Arabic connectives and the Arabic lexical stratum of ordinary Persian are excluded by the frozen definition.

The primary count is the 15 loci in حکایات ۱–۳۱, the part of the chapter this arm has now translated. That is a departure from the frozen design, declared in coding.md §Departure: the alignment work for the six later loci was not finished in the session, and they are reported as a partial supplement rather than dropped. 2 of the 60 primary cells could not be aligned (Ross at b022, b201) — 3.3%, inside the frozen failure criterion of 25%.

The four hands are A-gulistan-hands's: Gladwin 1806, Ross 1823, Eastwick 1852, Arnold 1899, public domain, read from Internet Archive scans. One instrument change, a gate rather than a unit: _djvu.txt throws away font information and italic is one of the four marking signals, so the _abbyy.gz layer was parsed instead (abbyy_italics.py). Without it the study's main finding is invisible.

2. The result

The registered prediction was that unframed loci would be marked at 0.35 or below. They are marked more than half the time. V5's premise — that where the author gives no signal the translator has none to give — is not the tradition's premise. Two of the four hands supply one, systematically.

Four hands, four policies, no two alike. Gladwin never marks; Eastwick marks almost always; Ross marks and glosses twice over; Arnold deletes.

Three things the count did not have to produce

1. Eastwick marks at the hemistich. At b236 the couplet is «زینهار از قرینِ بد، زِنهار! || وَ قِنا رَبَّنا عَذابَ النّار» — a Persian half-line and a Qur'anic half-line that rhyme. Eastwick prints:

From a vixen wife protect us well, Save us, O God ! from the pains of hell.

The Persian half roman, the Arabic half italic, inside one couplet. The prediction that no hand would mark this locus is false, and the reason it was made — that carrying it needs both a rhyme and a language change — turns out to be the wrong test: Eastwick does not carry the rhyme, he carries the language change, which is the part the arm had written off as unreachable.

2. Whether a hand keeps Sa'di's doubling is decided by whether it marks. At b218 Sa'di quotes a hadith in Arabic and then translates it into Persian himself in the same sentence, so the English would say one thing twice. All four hands are consistent, and they split two and two:

A hand that will not mark the Arabic cannot keep the repetition, because unmarked it is not a repetition, it is a stammer. This was decided independently in the translation limb the same morning, before any of these books was opened (R05-v1 log D30, register V15), on exactly that reasoning; the four hands are unanimous that the two decisions go together.

3. The one locus every hand leaves plain. b060, «مُشاهَدَةُ الاَبْرارَ بَیْنَ التَّجَلّی وَ اْلاِسْتِتارِ» — the vision of the pious is between manifestation and concealment. It is the only unframed Arabic in the fifteen that is not a quotation set beside the argument but a definition continuing it, and all four hands absorb it into the surrounding sentence. Ross even lets it run on into the Persian without a break. internal-judgment-only, on n = 1.

3. What the registered predictions did

Three of four predictions came out partly or wholly wrong, and they were wrong in one direction: the study underestimated how much these translators intervene.

4. What this changes

For the arm. V5 is not amended — it is a declared policy and the arm keeps it — but the sentence inside it that says the loss "cannot survive" is too strong and is corrected: the language change can be signalled, half the tradition signals it, and what the arm is doing is declining a repair that is available, not recording an impossibility. That is a different claim and a smaller one. V15 (a gloss counts as a frame) is confirmed from outside, unanimously, at the one locus that tests it.

For the framework. framework/v0.2 §7.45, written from this page: where a source changes language and the target cannot, the translator's real choice is not whether to mark but whether to keep — because an unmarked foreign quotation that the author himself glosses becomes a repetition with no cause, and every hand that refused to mark it deleted it. Evidenced on one language pair, one book, four hands, fifteen loci.

5. Limits

• Four books, one chapter, one work. Nothing here is about translators in general, or about Persian-to-English in general. It is about what four named books do at fifteen named places.
• The codings are the lead's reading of uncorrected OCR, unanchored. Every marked cell is quoted verbatim in coding.md so the reading is checkable, and no coding turns on a garbled character — but no second reader has seen them. internal-judgment-only.
• Gladwin's zero is partly typographic. His 1806 text uses italic only for running heads and three editorial asides; he had no italic device in the body. His fifteen PLAIN cells therefore under-report his intent as much as they report it. He did have quotation marks, and used them at one locus of fifteen, which is the part of his zero that is a choice.
• Arnold's one unframed present cell makes his unframed rate (0/1) uninformative. What is informative about Arnold is the dropping, and that is unambiguous: 10 of 12.
• Two Ross cells could not be aligned and are excluded rather than guessed.
• The primary was narrowed from 21 loci to 15 after the design was frozen. The narrowing is by position in the chapter, not by outcome, and the excluded loci's partial codings — which point the same way — are printed in coding.md §Supplement so the choice can be second-guessed.
• SAY fired nowhere unambiguously. No hand wrote in Arabic where Sa'di had not. The two amplifications that come closest — Gladwin's this verse of the Koran and Ross's this text of the Coran — unpack a frame Sa'di supplied (آیت already denotes a Qur'anic verse), so both are coded SAY-SOURCE and excluded. The marking, where it happens, is typographic rather than verbal, which the design did not predict either way and which is the reason the ABBYY layer had to be fetched at all.