Repository path: workshop/experiments/E-20260812g-foreign-italic/design.md · rendered 2026-09-09

E-20260812g — the foreign-marking italic, and whether the target language decides its fate

Frozen before any Russian correspondence call, retention or licence figure existed. What was already known when these predictions were written is stated in §2.4 rather than claimed away.

1. The question

RS-20260812b counted, on seven tales of Poe in three witnesses, what happens to a source's emphasis marks. One row of that census is unlike the others:

The result page read a mechanism off the texts, by inspection rather than by test: Poe's _ennuyé_, _écarté_, _chapeau-de-bras_ are italicised because they are foreign to English. In Baudelaire's French they are not foreign, so the mark has nothing to do. In Japanese they become katakana, and katakana already says "foreign", so the function moves to the writing system.

If that reading is right, then the mark's fate is a property of the target language, not of the translator — and the way to test it is a third target that splits, where some of Poe's foreign words are ordinary borrowed words and some are still foreign. Russian at the turn of the century is such a target: парвеню, пируэт, барельеф, экарте were Russian words; peine forte et dure, er lasst sich nicht lesen, chapeau-de-bras were not.

What this unit teaches about translating literature (the subject rule, wiki/tracks.md): when a translator may drop a source's emphasis mark without losing anything, and when dropping it is a loss — decided by a property of the target that the translator does not choose.

2. Materials

2.1 The witnesses

The tale set is E-20260812b's, unchanged and not re-chosen: the seven tales of Baudelaire's Nouvelles Histoires extraordinaires that Sasaki also translated and Aozora carries. Balmont's Собрание сочинений Эдгара По (Скорпион, Moscow, 1901–1912) covers all seven, so the Russian witness is one hand over the identical source-span population — 74 source spans, 68 italic and 6 small-capital, in admitted prose.

build_ru.py writes materials/<slug>.ru.txt, one paragraph per line, every italic rewritten _span_, {{разрядка}} and <i> folded into the same channel, {{lang}} not folded (it is a language tag, not an emphasis device), refs and the {{Отексте}} header stripped.

2.2 What the transcription gives, and what it withholds

Four of the seven Wikisource pages name no copy-text, and none of the seven is scan-backed. That is a harder provenance floor than E-20260812b had on its French and Japanese witnesses, and it is the reason for F3.

Erratum and corpus repair, 2026-08-12, before any correspondence call was made. Two things in the table above are wrong and both were found by a reference cell, not by inspection: the REF-B row of contamination.py compares Balmont against himself on two different tales and returned a 64-token identical run, which no two translations of different texts can have.

1. The extraction was including page furniture. Wikisource category footers, the editor's notes section — which glosses Poe's foreign phrases in italics, i.e. exactly the spans this run counts — {{right|}} epigraph blocks, and the source-edition line. build_ru.py now cuts all four (body()), and REF-B falls to a 5-token run with zero shared 7-grams, which is what an honest floor looks like. The span counts change: pit-pendulum 40 → 31, silence 4 → 3, total 128 → 118. The other five tales are unchanged, and belfry stays at 0, so F3 stands exactly as registered.
2. Six of the seven pages do name a copy-text after all; the first scan of the headers missed field names spelled with spaces and belfry's, which sits in a <u> line rather than a header field. Corrected: black-cat / usher / silence — Скорпион 1901, vol. I; william-wilson — Собр. соч. в 7 т., vol. VII, 1906, routed through az.lib.ru, a re-keyed e-text rather than a scan; pit-pendulum — Собр. соч. в 7 т., 1906; belfry — Скорпион 1911, vol. III; man-of-crowd — Скорпион 1906. One hand, but not one edition, and no scan behind any of them. The provenance limit is milder than §2.2 said and is still real.

Neither correction touches a prediction, and both were made before any span was read.

2.3 The lead as a fourth hand — the translation limb

The lead translates into Russian, under R04 (lead close, source-only), every paragraph carrying a FOREIGN span in the two tales that have the most of them — «William Wilson» ¶10, 11, 12, 36, 38, 39 and «The Devil in the Belfry» ¶19, 21. That rule is fixed by the FOREIGN counts of RS-20260812b §4.2, which existed before this design. 1,533 source words; ~~13~~ 14 of the 25 FOREIGN spans; the project's first translation into Russian.

Erratum, 2026-08-12, entered after the freeze commit 2217439 and before any figure existed. The frozen text said 13. The correct count is 14 — «William Wilson» carries 8 FOREIGN spans and «The Devil in the Belfry» 6. Arithmetic in the design, not a change of scope: the paragraph list is unchanged and was already exhaustive. Recorded here rather than silently corrected.

The point of the limb is not a fourth number. It is that this hand can be asked why, and Balmont cannot: the translator's log records, per mark, what the alternatives were and what decided it, written at translation time and frozen before any Russian span was read.

2.4 What was known when this was frozen — stated, not claimed away

• The Russian corpus had been fetched and counted: §2.2's paragraph and span counts, including belfry's zero, were on the screen when these predictions were written. F3 exists because of it and is registered here rather than discovered later.
• No Russian prose had been read. Every diagnostic run before this freeze printed counts, or printed markup with every Cyrillic character masked to ·. This is R04 §1 discipline for the translation limb and it is also the honest condition for a preregistration.
• The lead knows RS-20260812b's figures for French and Japanese by heart; they are in §1. The predictions below are therefore exposure-conditioned, exactly as that run's were (its critic finding 6), and P2's bar is Baudelaire's own 0.320.

3. The classification the primary rests on

N — for each of the 25 FOREIGN spans: in Russian literary usage around 1900, is this an ordinary borrowed Russian word (NATURALISED) or still a foreign word (STILL-FOREIGN)?

• The lead classifies all 25 and seals the classification before any Russian span is read (n_lead.json, committed with this design).
• Two non-Anthropic panel seats classify the same 25 blind: given the English span and its English sentence, told the target is Russian of about 1900, shown no Russian text, no predictions, and not told what the classification is for.
• Majority of three rules. The lead is one vote, not the arbiter.

4. Predictions, registered

P6 is a translator predicting his own behaviour and is not evidence of the mechanism; it is registered so the log can be wrong in public, and P6c replicates RS-20260812b §5's P5c.

5. Failure criteria, registered

• F1 — if fewer than 20 of the 25 FOREIGN spans can be located in Balmont's text, P1 is not computed and the run is reported as a location failure.
• F2 — if P3 fails, P1 and P2 are uninterpretable and reported as such: a hand that does not carry emphasis at all cannot inform a claim about which emphasis it carries.
• F3 — a tale whose Russian witness carries zero emphasis spans cannot distinguish "the translator dropped them" from "the transcription dropped them". It is not excluded; every headline figure is reported twice, with and without it. Belfry is that tale, it holds 6 of the 25 FOREIGN spans, and including it works against P1 and P2 — so the all-seven figure is the conservative one and is the primary.
• F4 — if the blind classification differs from the lead's on more than 8 of the 25 spans, N is reported as unstable and P1 is recomputed on the two blind seats' majority alone.
• F5 — if the pre-run critic returns NEEDS-REDESIGN, every finding is answered in writing before any figure is reported, as E-20260812b did.

6. Procedure, in order

1. build_ru.py — fetch, counts only. Done before this freeze (§2.4).
2. Freeze this design and n_lead.json. Commit is the freeze.
3. Translate the eight paragraphs into Russian from Poe alone; write the log at translation time; commit. Commit is the freeze.
4. dependence_check.py, by script, lead-RU against Balmont-RU: report the longest common run and the shared 7-gram count, and set contamination: from the measurement.
5. Pre-run critic — one non-Anthropic seat, the frozen design, NEEDS-REDESIGN allowed.
6. Blind classification N — two non-Anthropic seats.
7. Census: locate all 74 source spans in Balmont by reading, code CARRIED/DROPPED, and count every emphasis span in his admitted prose for the reader-side denominator.
8. verify.py — recompute every reported number by an independent path, plus planted mutations.

7. What this design cannot answer

• Nothing about readers. There is no jury in this run and no seat is asked whether a mark changes anything. Tier D is NOT PASSED; nothing here is a jury result.
• Nothing about Balmont's intentions. The estimand is correspondence between two named transcriptions, and for four of the seven tales the Russian transcription does not name its copy-text. "Balmont dropped" means "the Russian witness does not mark where the English witness does".
• Compensation is measured only where it is visible as script. P5 counts one compensation route. A translator who renders an italic as a lexical intensifier still scores as a drop.
• The lead is not an independent hand and its Russian is unassessed for quality. §2.3's limb is a decision record, not a comparison of translations.