Repository path: workshop/experiments/E-20260812b-emphasis-carriage/design.md · rendered 2026-09-09

E-20260812b — emphasis carriage: three witnesses, one author, two hands

Frozen 2026-08-12 (S164) before any emphasis statistic was computed. Exposures the lead already had when this was written are declared in §8 and are not hidden behind the freeze.

1. Question

When a reader of a translation sees a word set in emphasis type, was it emphasised in the source? When the source emphasises a word, does the translation's reader see it?

RS-20260811c established the demand side: an added italic fooled every seat on every segment about how the original is written, at accuracy 0.0000, the strongest of four planted lures. RS-20260812 established the supply side for !, ?, — and … and could not touch emphasis, because the transcriptions it had did not mark it. This run supplies the missing side.

2. Materials — three witnesses, and why each is readable

All three are public domain, and the translators' rights are cleared separately from the author's: Baudelaire d. 1867 (PD in life+70 since 1938; French publication 1857/1884), Sasaki d. 1943 (PD in life+70 since 2014; the Aozora card is published under Aozora's PD terms).

The tale set is fixed by a rule that cannot see an emphasis mark: every tale in Baudelaire's Nouvelles Histoires extraordinaires (1857) that Sasaki also translated and that Aozora Bunko carries. Membership of one 1857 collection, intersected with one translator's output. Seven tales result:

black-cat (Le Chat noir / 黒猫) · usher (La Chute de la maison Usher / アッシャー家の崩壊) · william-wilson (William Wilson / ウィリアム・ウィルスン) · pit-pendulum (Le Puits et le pendule / 落穴と振子) · belfry (Le Diable dans le beffroi / 鐘楼の悪魔) · silence (Silence / 沈黙) · man-of-crowd (L'Homme des foules / 群集の人).

Why the Japanese is read from raw XHTML. Method note (bcg) was written on exactly this material: tools/fetch_aozora.py discards the annotation markup, 傍点 lives in it, and a text that has lost every mark and kept every character reads as complete. build_corpus.py reads the XHTML and rewrites the sesame spans before any tag is stripped.

Everything is normalised to one convention, _span_, in all three languages, so that one counter runs over all three witnesses and no witness gets its own code path.

3. Procedure

1. build_corpus.py — extract, per tale, EN/FR/JA one paragraph per line, plus an enkind file typing each EN block P prose or V verse-or-epigraph. The typing rule is the copy-text's own indentation: in these Gutenberg volumes every prose line is indented exactly six spaces and every line of inset verse, verse numeral or epigraph is not. It reads the source alone; no target text can move it.
2. align.py — length-based dynamic-programming alignment (Gale–Church shape), EN as pivot, run once per (tale, hand). Links allowed: 1-1, 1-2, 2-1, 1-0, 0-1, 2-2, with a fixed penalty on each non-1-1 link. The length function deletes the _ delimiters before measuring, so the aligner is blind to the marks it is being used to count. The only free quantity is the length ratio, estimated from the two texts' own totals.
3. Admission — an aligned unit is admitted iff its EN side is non-empty and every EN member is kind P. Verse, epigraphs and the inset "Haunted Palace" fall out by the rule of step 1, not by hand. Target blocks that align to nothing are reported with their mark count and excluded.
4. Source-side type coding — every EN emphasis span is typed once, by the lead, into STRESS (emphasis on an English word or phrase), FOREIGN (a word or phrase not English), TITLE (a named work, ship, publication or institution) or OTHER. This is a property of the source alone and cannot be pulled by either target; it is nonetheless single-coded and that is declared. The coding is written to en_types.json before any target-side statistic is computed, and the file is committed in that state.
5. Counting, per hand H ∈ {FR, JA} over admitted units u: - retention, site-level: for each EN span, 1 if the aligned H block contains at least one H span, else 0. Reported overall and by type. - licence, mark-level: for each H span, 1 if the aligned EN block contains at least one EN span, else 0. licensed(H) is the mean — the fraction of the marks that hand's reader sees which are licensed by a source mark in the same place. This is the statistic RS-20260812 put at 0.9425 for punctuation and could not compute for emphasis. - added = H spans with no EN span in the aligned block; omitted = EN spans with no H span in the aligned block.
6. Grain. Everything is at paragraph grain, exactly as RS-20260812. A mark that moves within its paragraph counts as carried. Retention and licence are upper bounds; added and omitted are lower bounds. Site-level correspondence is worse than reported by an unmeasured amount, and no claim here may be stated at site grain.

4. Translation limb

T-black-cat-R04-v3 completes the lead's Japanese rendering of "The Black Cat": paragraphs 1–3, 10–13 and 21–30 of the stored source, 2,146 words, joining R04-v1 (¶14–20, S038) and R04-v2 (¶4–9, S063). The story is then rendered whole in one hand.

The lead is NOT an independent second Japanese hand and no figure here treats it as one. The standing gate was run before this design was written, not inside it: R04-v1 against Sasaki gives a longest common run of 16 characters and R04-v2 gives 17, against reference cells at 8 and 9 — the lead's Japanese is measurably closer to Sasaki than Sasaki is to himself on other content. contamination: high is declared on the new artifact on those measured numbers.

What the limb is for is the one thing no published hand in the corpus can supply: a written record of the decision. Baudelaire and Sasaki left no account of a single emphasis mark. The translator's log records, for each of Poe's marks in the new span, what was done — and carries a sealed prediction, written after translating and before any count, of what the lead's own 傍点 did across the whole story.

5. Registered predictions

Census.

• P1 — Baudelaire's reader sees more unlicensed emphasis than Sasaki's. licensed(FR) < licensed(JA). Reason: French italics do several jobs at once — emphasis, foreign words, titles, ship names — and Poe writing in English had no reason to mark the English realia that a French page must set apart, so Baudelaire acquires marks the source has no place for. Japanese 傍点 does one job, and Japanese has other machinery (katakana, 「」) for the rest.
• P2 — licensed(FR) < 0.85. A quantitative form of P1, chosen so that it can fail while P1 holds.
• P3 — FOREIGN retention is lower than STRESS retention, in both hands. Reason: a French phrase italicised as foreign in English is not foreign in French, and Japanese does not use 傍点 for foreignness at all.
• P4 — STRESS retention exceeds 0.60 in both hands. The claim that a published translator simply drops the emphasis channel is the one this run can refute, and predicts it will.

Translation limb.

• P5 — stated in the translator's log, sealed in a commit before any count of the lead's Japanese exists. Its content is not repeated here so that this file cannot be edited to match.

6. Failure criteria — what stops a number being reported

• G1 alignment. For a (tale, hand) cell, if fewer than 60% of admitted units are 1-1 links, the cell is excluded from the pooled figures and reported separately with its profile. Alignment quality is a property of the two texts and is computed before any mark is counted.
• G2 witness. If a hand shows zero emphasis marks in three or more of the seven tales, that hand is a transcription artefact rather than a translator and no licence figure is reported for it. (RS-20260812 excluded Garnett on the sibling rule.)
• G3 mutation. The verifier must recompute every reported number by an independent path and must fail on at least five planted mutations of the stored corpus. A verifier that passes a corrupted body has checked nothing.
• G4 transcription, declared not gated. All three witnesses are modern transcriptions of particular printings — Gutenberg's of an 1890s collected edition, Wikisource's of Quantin 1884, Aozora's of a postwar 底本. Any of these marks may be a compositor's or an editor's rather than the author's or the translator's, and none of it is recoverable from the files. This is not gated because it cannot be: it is the reader's situation too. It bounds every claim about intention and bounds nothing about what a reader of these printings sees.

7. What this run cannot license

• Nothing about human readers. The demand side is RS-20260811c's, and that was prompted LLM seats with Tier D NOT PASSED.
• Nothing at site grain (§3.6).
• Nothing about the lead as an independent hand (§4).
• Nothing about compensation: a translator who renders Poe's italic stress as a lexical intensifier scores as an omission here. The census counts marks, not emphasis.

8. Declared exposure — what the lead had already seen when this was frozen

Written out rather than claimed away, because a freeze that hides what preceded it is worthless.

1. Whole-text totals, from building the corpus: "The Black Cat" EN 19 italic runs and FR 20; Sasaki 傍点 span totals per tale (black-cat 20, usher 15, william-wilson 44, pit-pendulum 27, belfry 9, silence 4, man-of-crowd 3). No per-paragraph, per-site or aligned figure was computed, and no correspondence, retention or licence quantity exists yet. P1–P4 are about correspondence, which totals do not determine.
2. The italic word-list of "The Black Cat" was printed while testing the extractor.
3. Method note (bcg), S038, states that on "The Black Cat" Sasaki "marks six of seven and adds one". That is a prior on one tale and on a differently-drawn site list, and it is the reason P4's bar is set at 0.60 rather than at a coin.
4. The lead's own Japanese was not examined for 傍点 at any point before P5 was sealed.