Repository path: framework/v0.3/entries/HB-change-of-language.md · rendered 2026-09-09

When the source changes language and the target cannot

Standing. Four pairs, three authors, twelve published hands 1806–1923 and 1879, 240 English cells (58 of the 60 Persian cells aligned) plus one French hand checked on 50 spans, all coded by the lead from scans and electronic texts (one study with a blind second coder), so internal-judgment-only; every number is a count of what a hand did on a named page, never a judgment that a hand did better. Tier D is NOT PASSED: nothing here rests on a jury. Consolidated 2026-09-04 (S244) from v0.2 §7.45, §7.50 and §7.52; the exemplar entry of v0.3. Item 12 and the RU→FR pair added 2026-09-06 (S251), evidenced after consolidation, not from v0.2.

1. The problem

An author writes in one language and, at some places, another: Sa'di's Persian quotes Arabic scripture, maxims and verse; Dante's Italian quotes Latin — poets, scripture, and the speeches of figures in his own book; Tolstoy's Russian aristocrats simply speak French, and Tolstoy prints his own Russian gloss at the foot of the page. The source reader sees the language change. An English page cannot change language in the same way: whatever the translator does at each such place — keep the foreign words, mark them, gloss them, english them, or cut them — is a decision the author never had to make, and it recurs at every locus. The evidence below is about what published translators actually did at 80 such loci in three books, and what the option set turns out to depend on.

2. What published translators do

The regularities, each with what it was measured on:

1. Quotation and code-switch are different problems, not different degrees. Dante's hands keep the Latin at 70 of 84 cells (0.83); Tolstoy's keep the French at 15 of 96 (0.156). The readerships are the same Victorians. What differs is that a quotation is somebody else's words, bounded and attributable, and a code-switch is the character's own voice: a translator leaves another writer's sentence standing in its language and does not leave his character's speech standing in one. (§7.52.1; RU→EN against IT→EN.)
2. Where the reader can read the embedded language, nobody deletes. Zero deletions in 84 Latin cells and zero in 96 French cells. (§7.50.1, §7.52.2; IT→EN, RU→EN.) Where the reader cannot — Arabic for an English reader — a hand that will not mark the foreign line cuts it and prints the author's own gloss alone, four of four such cells (§7.45.3; FA→EN).
3. Marking is a contrast with the surrounding type, and italic is only its commonest form. Among Dante's retained Latin, italic accounts for 62 of 70 cells; but three hands set Dante's analytic paragraphs in italic as a body, and there the Latin is set roman. Measured as type-against-surrounding-type, 69 of 70 retained cells are distinguished; the one exception is italic inside italic. (§7.50.2; IT→EN.) A study that codes "italic or not" mis-scores every locus whose surroundings are already italic.
4. Half a tradition supplies a signal the author did not. At the Persian loci Sa'di does not frame, the pooled marking rate is 19 of 35 (Eastwick 11 of 12, Ross 8 of 10); at the loci he frames, 9 of 12. The marking is typographic, not verbal — no hand wrote in Arabic where the Persian had not. (§7.45.1; FA→EN, one chapter.) A hand's silence is evidence of a choice only where the hand had a device: Gladwin marks at 0 of 12, and his book uses italic only for running heads. (§7.45.4.)
5. The signal marks the language, not the status of what is said. Across the two marking Persian hands, the marking rate is 1.000 on fourteen ornamental cells, 0.929 on fourteen scriptural, 0.800 on ten gnomic; Ross italicises an Arabic couplet about pomegranate blossom and leaves a Qur'anic hemistich plain. (§7.45.5; FA→EN.) On Dante the registered form of this claim failed for two hands by one cell each — the italic-inside-italic cell of item 3 — and survives only on the contrast measure, which was chosen after the numbers were seen; what is safe to say there is that no hand marks scripture and leaves dramatic speech unmarked, or the reverse. (§7.50.3.)
6. Glossing, by contrast, does track function — where the quotation is somebody speaking. Dante's hands supply an English rendering beside the retained Latin at 21 of 21 cells where a figure in the book speaks Latin, and at 8 of 27 where Dante cites a learned authority (Fisher one-tailed P = 1.9 × 10⁻⁷; the contrast was chosen after seeing the table and is descriptive). (§7.50.4; IT→EN.)
7. At a code-switch, englishing is the gloss, moved off the foot of the page and into the text. Tolstoy's dual text — French in the body, his own Russian below — is collapsed into one English text at 91 of 96 cells; where the hands do leave French standing, they print nothing beside it: one glossed retention in 96 cells (a registered prediction of ≥ 0.80 failed on every hand). (§7.52.3; RU→EN.)
8. The supplied signal changes register with the choice: typographic where the language is kept, verbal where it is collapsed. A collapsed switch leaves nothing to italicise, so a hand who wants the reader to know a language changed says so in words — Maude's said she in French, still in French; Bell's and in French — four verbal signals against one gloss; Garnett gives no indication anywhere in the two chapters that a second language is spoken. (§7.52.4; RU→EN.) Whether these hands italicise the French they keep could not be measured: the electronic witnesses carry no italics. (§7.52.6.)
9. A switch inside a switch is the first casualty. At the five loci where Tolstoy plants Russian words inside a French sentence, every hand englishes the whole; the inner switch survives in 1 of 12 available cells — Maude's quotation marks around 'faithful slave' — and he does not repeat the device eighteen loci later at the same formula. (§7.52.5; RU→EN.)
10. The unattributable quotation is the one a tradition will not leave standing. Of Sa'di's 21 Arabic loci, exactly one is deleted by all four hands: an unattributed, unframed Qur'anic rhetorical question. Of Dante's 21, the one classical authority the englishing hand puts into English is the only quotation in the book with no author's name on it (she keeps the other 6 of 6; Fisher P = 0.0005). (§7.45.6, §7.50.5; FA→EN, IT→EN.)
11. An author's own printed language policy does not bind his translators. Dante says inside the book that he will write only in the vernacular; all four hands print the sentence and all four break it, and de Mey prints "the words cited are all in Latin" four sentences after printing those words in English. (§7.50.6; IT→EN.)
12. Where the target coincides with the embedded language, verbatim carry is not the default, even though it costs nothing. Translating Tolstoy's own French into French, the one published hand measured (Paskévitch 1879) rewrites nearly everything Tolstoy wrote in French rather than reusing his words — of 50 checked runs of two or more words, 14 survive verbatim and most of those are short conventional tags (ma tante, l'abbé Morio) that two independent hands would likely share regardless of copying. The one true exception is a quoted written document (Tolstoy's invitation letter), kept nearly word for word — she still edits one clause (the hour). The operative boundary is not "was this French in the source" but is this quoted as a document, the same test item 1 uses to separate quotation from code-switch, now appearing inside a single translator's own practice. First evidence added after consolidation, not from v0.2. (RS-20260906-french-target-voina-i-mir; RU→FR, n = 1 hand.)

3. What this project's own practice found

The lead rendered the same three problems under a written rule — do not supply a signal the author did not give — and the rule put it outside the entire published tradition twice (internal-judgment-only throughout; the lead's renderings are labelled hands, never independent witnesses — contamination DEPENDENT? against three of four Dante hands and all three Tolstoy hands).

• Dante (T-vita-nuova-R04-v1, XXIV–XXVI, 1,715 Italian words): at the seven Latin loci in the span, kept 7 of 7, typographically distinguished 0 of 7, glossed 0 of 7; the four published hands keep 7, 7, 7 and 6 and distinguish every one they keep. The page tells the reader nothing at the change of language, which is what the source does, and no published hand does it.
• Tolstoy (T-voina-i-mir-R04-v1, I.i.1–2, 2,780 Russian words): the dual text is carried — French in the body, Tolstoy's own footnoted glosses englished at the foot — and the Russian planted inside the French becomes English inside the French (des estates, de la famille…; vous n'êtes plus my faithful slave), so the inner switch survives as a switch. Cost: the reader who cannot read French reads the scene through footnotes, exactly as Tolstoy's non-Francophone Russian reader did. One place the rule broke: the раб / рап misspelling joke, where the lead's invented slafe does a job Russian orthography does for free — Maude's slave-slafe with an f is the better-recorded solution (journal 2026-09-04; an observation, not a score).
• Tolstoy, into French (T-voina-i-mir-fr-R04-v1, same span, same 2,780 Russian words): the target coincides with the embedded language, so the code-switch itself has no move — the policy adopted was italic for Tolstoy's French verbatim, roman for translated Russian, Cyrillic-plus- gloss for the five loci where Tolstoy plants Russian inside French, which are now the only genuinely foreign material left in the text and, at two of the five, require no translation at all (the source already alternates French/Russian in the shape a French target needs). Measured against Paskévitch 1879, the one published hand takes the opposite approach at every point except a quoted document: she rewrites nearly all of Tolstoy's own French rather than reusing it, and assimilates the раб/рап joke into an invented French misspelling (esclafes) rather than leaving the Russian to stand, an option this session's rendering used and hers did not (RS-20260906-french-target-voina-i-mir).
• Sa'di (T-gulistan-bab2-R05-v1, spans B–E under a binding register): Arabic kept plain at every unframed locus; b346, span E, is the chapter's one framed Arabic locus («موافق قرآن است», "it agrees with the Qur'an") and it is also followed by a Persian couplet that restates its point — the shape span B–D called a gloss (V15) when it occurred unframed. Framing and glossing turn out to answer two different questions (does the author name the source; does he then say it again in his own words), not one, and a locus can carry either, both, or neither: b346 is marked because it is framed, regardless of the couplet that follows it. This is a clarification of item 2 below, not a correction — the two rules were never actually in tension, only stated close enough together to look as if they might be.

4. The options

At each locus the moves are keep (marked or unmarked; glossed or not), english (collapse the change of language into the target), and cut (print the author's gloss alone, or nothing). The evidence says the option set is not fixed; it depends on two things:

• Whether the target reader can read the embedded language. Latin and French for Victorian English readers: the third option is english, deletion never occurs, and the trade runs between marking and translating. Arabic: the third option is cut, and the decision to mark decides the decision to keep. (items 2, 4.)
• Whether the passage is a quotation or the speaker's own voice. Quotation: keep, marked by contrast, glossed where somebody speaks it. Code-switch: english, and signal in words if the switch matters — the dual text is the alternative no published hand took. (items 1, 6, 7, 8.)

Two costs are on record. A kept, glossed quotation says one thing twice unless the first is marked (item 2). A collapsed switch loses the second-order switch (item 9) and everything the switch was doing in the scene, at the gain of a page that reads without apparatus.

5. Guidance

For a translator

1. Enumerate every change of language from the source before you open any other translation, and classify each locus twice: quotation or code-switch; and, within quotation, scripture, authority, maxim, or a figure speaking. The rest of this entry is decided by that classification. — evidenced (FA→EN, IT→EN, RU→EN).
2. Do not delete what was said where your reader could read it. Where your reader cannot, and the author glosses the quotation himself, decide whether you will mark before you decide whether to keep: if you mark, keep the author's doubling; if you will not mark, the tradition cuts the foreign line and prints the gloss — printing it plain is the option no hand took, and it is what the source does. Framing and glossing are independent checks, not one rule with two names: ask both of every locus — does the author name the source (frame), and does he then restate it in his own words (gloss) — because a locus can answer yes to either, both, or neither, and the frame answer settles marking on its own, whatever the gloss answer is. — evidenced (FA→EN for the cut; IT→EN, RU→EN for no-deletion).
3. Mark a kept quotation by contrast with the type around it — italic in roman, roman in italic — and mark the language, not the status: a hand that italicises scripture and not verse tells the reader something the author did not. — evidenced (FA→EN, IT→EN).
4. Gloss a kept quotation where somebody in the text is speaking it; leave a cited authority standing if that is what the tradition of your target does — this is what four Dante hands did at 21 of 21 and 8 of 27, and it is a description of practice, not a measured preference. — evidenced (IT→EN), descriptive.
5. At a code-switch, choose between carrying the author's dual text and collapsing it. If you collapse, italics are gone: the only device left for the change of language is saying so (she said in French), and the second-order switch will be lost unless you mark it inside the target (quotation marks did it once in 96 cells). If you carry, expect to supply the gloss yourself; published hands supplied one at 1 of 15 retentions. — evidenced (RU→EN).
6. Decide the unattributable quotation deliberately. It is the one locus two traditions singled out — cut by all four Persian hands, englished by the one Dante hand who kept every named authority. — evidenced (FA→EN, IT→EN).
7. Do not expect the author's printed language policy to settle anything; four hands broke Dante's on the page where they printed it. — evidenced (IT→EN).

For a pipeline

1. Enumerate: detect every run in a second script or language in the source text; freeze the locus list with its extent, before any target text exists. — untested.
2. Classify each locus (quotation / code-switch; function; attributed or not), by a model pass with the definitions from §1, and hold the classification frozen. — untested.
3. Set the policy parameter: reader can read the embedded language (from the declared purpose of the translation, wiki/goodness-senses.md §declared-purpose) and the move per class from §5 above (keep-marked-glossed / keep-marked / english-with-verbal-signal / carry-dual-text). — untested.
4. Render under the policy; at a collapsed switch, emit the verbal signal where the source marks the switch as consequential (a locus the classifier tagged as a turn in the scene). — untested.
5. Check: no locus deleted; every kept run distinguished from its surrounding type; every kept locus in a speaking class glossed; every second-order switch either carried or marked; report the per-class counts against the policy. — untested.

Human entry points. Step 3 (the policy) is a purpose decision a person should make; step 2's attributed/unattributed calls and any locus where the classifier and the policy conflict are the places to route to a person. Without a person, step 3 defaults to keep-marked for quotation and english-with-verbal-signal for code-switch — the majority practice in the evidence.

Application. W3 step 1 (Gulistan span E, S244) applied items 2–3 at b346, the chapter's one framed Arabic locus: marked (italicised) because framed, per item 3, with item 2's frame/gloss distinction confirmed rather than complicated by the couplet that follows it (D63, workshop/translations/gulistan-bab2/register.md). No followability cost recorded — the decision was immediate once frame and gloss were asked as separate questions. W3 step 3 (War and Peace into French, S251) filled the one direction item 5 could not advise on — the target coincides with the embedded language — and found a fourth move (item 12): collapse everything except an actually- quoted document, and put any marking device on the translator's own coinages rather than on the author's original switch. Followability cost: real, on the one long stretch of the chapter that is entirely Tolstoy's Russian with no French to anchor register (Anna Pávlovna's tirade on Austria, RS-20260906-french-target-voina-i-mir §7 of the translation's own log).

6. Not evidenced, and open

• Eleven hands are eleven books, not eleven draws; two Tolstoy hands measure as dependent on each other, and nothing separates a period convention (1806–1923) from a translator's decision. No hand after 1923 has been measured.
• One coder throughout, the lead, from uncorrected OCR; a blind second coder was bought only on Tolstoy (agreement 92 of 96, and it flipped one registered prediction on a lead coding error — that prediction, on switch type, is unresolved).
• The typographic dimension is missing on Tolstoy: no italics in either electronic witness.
• Nothing on what a mark, a gloss or a verbal signal does for a reader; the shown-versus-told finding (HB-telling-the-reader) was measured on a different problem.
• One published French hand, one lead French rendering — n = 1 on each side of item 12; not known whether Paskévitch's total-rewrite practice or the lead's dual-text-collapse-with-Cyrillic practice is more typical of a translator facing this cell.
• Open: whether a modern hand (post-1923) collapses a code-switch at the same rate as the Victorian and Edwardian hands measured here. Closed at S244: Sa'di's one framed locus, b346 — see §3 and item 2 above. Closed at S251: the French-target case (W3 step 3, Paskévitch 1879) — see §3 and item 12 above.

7. Sources consumed

• v0.2 §7.45 (S235) and §7.45.5–6 (S240); §7.50 (S241); §7.52 (S243), with their Evidenced on lines and scope amendments — all consolidated; nothing added after S243 lives in v0.2.
• RS-20260831-arabic-in-persian §5 (limits), RS-20260902-arabic-function, RS-20260903-latin-in-italian §9 (limits; two arguable codings), RS-20260904-french-in-russian §8 (limits); A-gulistan-hands; the coding files of E-20260903-latin-in-italian and E-20260904-french-in-russian (every cell quoted verbatim).
• RS-20260906-french-target-voina-i-mir §6 (limits), added 2026-09-06 (S251), post-consolidation: item 12 and the practice-note in §3 on T-voina-i-mir-fr-R04-v1.
• Withdrawn or scoped along the way: §7.45.3 (decide whether to mark before whether to keep) scoped to option sets whose third move is a cut, by §7.50.1; §7.45.5's registered form failed on Dante for two hands and survives only on the repaired contrast measure, §7.50.3; §7.45 and §7.50 as wholes scoped to quotation by §7.52; the S243 switch-type prediction unresolved, and the ≥ 0.80 glossed-retention prediction failed on every hand, §7.52.3.