Repository path: workshop/experiments/E-20260731c-futabatei-alignment/design.md · rendered 2026-09-09

E-20260731c — the dialogue-anchored alignment, and what Futabatei's rule costs in the language he wrote it in

ARM-nonlead-log step 2(a). Session S071, 2026-07-31. This design is frozen in git before the translation limb is written and before any paragraph of 「あいびき」 outside the exposure record below is read.

0. Why there is a second attempt

E-20260730g condition B asked whether 二葉亭四迷's 1888 「あいびき」 complies with the punctuation rule he stated in 1906, and voided on its own registered criterion F1: a monotone length-based alignment of 69 Russian paragraphs against 75 Japanese ones put only 51.4% of the Russian's words into 1:1 blocks, and hand-verification found the aligner wrong at three of its eight non-1:1 blocks. RS-20260730g §4 names the repair and ARM-nonlead-log step 2(a) carries it: anchor the alignment on dialogue structure — Russian speech paragraphs open with «—», Japanese ones with 「 — instead of on length, and measure at the aligned-block level rather than on the 1:1 subset.

And RS-20260730g §3 names the other half, against itself. It measured what the rule costs in English — J1 free, J2 nearly free, the whole residual one grammatical construction — and then wrote: "the difference is not evidence about Japanese." Condition D tried to reach Japanese with two generator models and a naturality rater, and the rater scored Futabatei's own prose at 2.25 of 5, which voided it. A compliance number for Futabatei is uninterpretable without knowing what an unforced translator into Japanese does with the same source. That is this session's translation limb, and it is free.

The wire, in one sentence. The translation limb supplies the unforced-Japanese baseline without which the study limb's compliance figure for Futabatei cannot be read as evidence about his rule rather than about the Japanese language.

1. Exposure record — read this before believing any blind claim here

The 「-initial heuristic that ARM-nonlead-log step 2(a) prescribes is confounded, and establishing that required reading. Futabatei writes transliterated proper names inside 「」 — 「ヴィクトル」, 「アクーリナ」 — so a paragraph that opens with 「 may be narration whose subject is a name. This was found by printing paragraph heads. Recorded exactly:

Nothing of ¶6–¶15 or ¶44–¶75 has been read by this session, and ¶44–¶75 is where the translation limb's span lands. The span was chosen on the exposure record and on nothing else — no punctuation count, no compliance figure, and no sentence of the corresponding Japanese, had been computed or read when it was chosen. The inherited S066 exposure of the first 30 characters of ¶73–¶75 stands and is not repaired by anything here.

The residual risk, stated rather than argued away. ¶16–¶43 of the Japanese corresponds to roughly RU ¶17–¶41, which is outside the translated span but inside the measured one. It cannot prime the translation limb. It can prime the aligner, and the aligner is hand-verified by the lead, which is why stage 3 exists.

2. Materials

Paragraph numbering. Artifacts and prose here use 1-indexed ¶ (¶1 = file index 0), matching T-svidanie-R06-v1 and T-svidanie-R09-v1. Code is 0-indexed.

The translated span is RU ¶42–¶69 (file indices 41–68), 669 Russian words, the story from «Прежде вы со мной не так говаривали» to the end. T-svidanie-R06-v1 / -R09-v1 (English, S066) hold ¶1–¶19; ¶20–¶41 is translated in neither language by the lead and that gap is deliberate — the span was chosen on exposure, not on contiguity.

Variant paragraphs. workshop/canon/svidanie/manifest.md establishes a noise floor of 5 commas and 1 sentence-ender between the two Russian witnesses, at file indices {0, 21, 22, 37, 49, 67}. These are excluded from every compliance statistic, exactly as in E-20260730g §3. Two of them, indices 49 and 67 (¶50 and ¶68), fall inside the translated span, so the translation limb's statistics are computed over 26 of its 28 paragraphs.

Counting definitions, frozen verbatim from E-20260730g/analysis/analyse.py so the figures are comparable (method note (r)):

RU_ENDER = r"[.!?…]+"      RU comma = ","
JA_ENDER = r"[。！？]|[…‥]+"  JA comma = "、"

A run of Russian ellipsis-plus-mark counts as one ender. This is inherited, it is arguable, and it is not changed here because changing it would make S066's English figures incomparable.

3. Stage 1 — the dialogue-anchored aligner (D1)

Frozen before it is run. Every constant below is a design commitment, not a fitted parameter.

RU type. ¶ is D iff it begins with — (em dash) after stripping whitespace; else N.

JA type — the repair's own repair. ¶ is D iff it begins with 「, except that it is N when all three hold of the first 「…」 span s and the character t immediately after its closing 」:

1. t ∈ {は, が, も, の, を, に, へ, で, や, と, ト} — s is a sentence constituent, not an utterance; and
2. s contains no character in 。！？、 — a name is not punctuated; and
3. len(s) ≤ 8 — a name is short.

A ¶ not beginning with 「 is N.

This rule is structural and is frozen now. It was written from having seen five instances of the name-quote pattern (¶19, ¶25, ¶31, ¶35, ¶40) and no instance of it failing; every misclassification it makes anywhere in the work will be found by stage 2 and stage 3 and reported as a defect of this rule, not silently repaired.

Alignment. Monotone dynamic programming over RU ¶0–68 against JA ¶1–75 (¶0, the translator's note, has no Russian counterpart and is excluded by construction and declared here rather than discovered). Permitted block shapes: 1:1, 1:2, 1:3, 2:1, 1:0, 0:1. Cost of a block:

length term   |ja_chars − ratio · ru_words| / max(1, (ratio·ru_words)**0.5) * 4
split penalty 25   for any block that is not 1:1
gap penalty   40   plus the length term, for 1:0 and 0:1
TYPE penalty  60   per RU paragraph in the block whose type differs from the
                   MAJORITY type of the JA paragraphs in the block

ratio = total JA non-whitespace characters ÷ total RU words, fitted on the whole work, as in S066.

The type penalty at 60 is more than twice the split penalty, so the alignment prefers to split or merge rather than to cross a dialogue boundary. That is the whole content of "dialogue-anchored" and it is the one number this design invents.

4. Stage 2 — hand-verification, and the F1 repair (D2)

Every non-1:1 block is hand-verified by the lead and recorded, as E-20260730g required and as note (bfa) says is what made the void verdict readable. In addition, twelve 1:1 blocks drawn by a seeded rule (random.Random(20260731), sampled from blocks not already non-1:1) are hand-verified, which S066 did not do.

Note (bfa) is discharged here. F1 conflated the instrument failed with the phenomenon is real, because both produce low 1:1 coverage. The criterion is split:

• F1a — instrument. If hand-verification finds the aligner wrong at more than 20% of non-1:1 blocks, or wrong at any of the twelve sampled 1:1 blocks, the alignment is not established and B is void.
• F1b — phenomenon. Low 1:1 coverage does not void anything. A translator who splits and merges paragraphs is a finding about the translator, and it is reported as a number with the block-type histogram beside it.

5. Stage 3 — the independent alignment check (D3) — the only API spend

The lead cannot both build the alignment and be its only verifier. Three panel seats (P1, P2, P3; reserve P5, declared now per note (bfc)) each receive the same frozen item list in one call: a Russian paragraph (or paragraphs) and five candidate Japanese paragraphs — the aligned one(s) plus the two before and two after in the Japanese — and are asked which of the five render the Russian. Multiple selection permitted; DECLINE permitted.

The prompt names no author, no date, no rule, no study. This is factual adjudication, which config/models.md §Instrument note (S015) licenses this panel for, and it makes no quality judgement of any kind.

Items: every non-1:1 block, plus the twelve seeded 1:1 blocks, capped at 30 items in source order if that exceeds 30.

• F2 — if the seats' majority answer disagrees with the lead's alignment at more than 25% of items, B is void and only the alignment and the translation limb are reported.

6. Stage 4 — condition B, Futabatei's compliance

Block level, not the 1:1 subset — this is the arm's own instruction. For each aligned block containing at least one RU and at least one JA paragraph, and containing no variant paragraph:

• J1 — does the sum of JA sentence-final marks equal the sum of RU sentence-enders?
• J2 — does the sum of JA 、 equal the sum of RU ,?

Reported: raw compliance, compliance restricted to 1:1 blocks (comparable to S066's void figures), and both against a 10,000-draw permutation null (random.Random(20260731)) in which the JA counts are shuffled across blocks. Mean absolute deviation is reported alongside the exact-match rate, because a rule that is nearly kept and a rule that is not kept at all give the same exact-match number.

7. Stage 5 — the translation limb (L)

Order, and it is not negotiable. L06 is written first, in one pass, from the Russian alone, with punctuation written as Japanese wants it and no attention paid to the Russian's counts, and is committed to git before L09 is begun. L09 is then written from the Russian under R09, with a numbered log citing the rules. The log is frozen before any count is computed (continue-prompt.md §5.4). Neither limb is judged for quality by anyone, and R09 §Roles says it makes no quality claim at all.

This is the project's first translation INTO Japanese as a filed T- artifact and the limitation that comes with it is stated at the top of both artifacts: the lead is not a native writer of Japanese and is certainly not a writer of 1888 言文一致 Japanese. What the limb bounds is what the rule costs this translator in modern Japanese — the within-translator delta from unforced to forced. It does not bound what it cost Futabatei, and no claim here will say it does.

Same statistics as stage 4, on the same definitions, over RU ¶42–¶69 minus the two variant paragraphs.

8. Stage 6 — contamination

Run after both limbs are committed, via tools/dependence_check_cjk.py, which requires reference cells:

• F4 — if a subject run fails to clear both reference floors, the limbs are declared contamination: none on the measurement. If it clears both, they are declared high and every figure that uses them is reported with the declaration attached, not withdrawn. Withdrawal is not on the table because the limbs' own comparison is within-translator, and priming toward Futabatei would move the lead's numbers toward compliance, which is against the interesting direction for PL4.

Ordering, stated as the limitation it is (note (bcd), the S063 form). The standing rule is measure contamination before selecting the material. It cannot be obeyed here: the material is fixed by the arm — Futabatei translated «Свидание» and no other work will do — and the comparator is the object of study. The prospective declaration, recorded now, before anything is translated: suspected, on the ground that the lead has read 28 paragraph-heads of the comparator and none of them in the translated span.

9. Predictions, registered

10. Failure criteria, registered

• F1a — aligner wrong at > 20% of non-1:1 blocks, or at any sampled 1:1 block → B void.
• F1b — low 1:1 coverage is reported, never voiding. (The note (bfa) repair.)
• F2 — independent seats disagree with the lead's alignment at > 25% of items → B void.
• F3 — if L09 cannot reach 100% on J1 and J2 at ≥ 90% of its paragraphs, PL3 is failed and the forced arm is reported as a partial ceiling, not as a ceiling.
• F4 — as §8.
• F5 — if any panel seat's answer arrives truncated or empty and its declared reserve also fails, that seat is dropped and the majority is taken over the remainder, with the drop stated. Note (b) has fired sixteen times; it will fire again.

11. What this experiment cannot show, written before it runs

1. One translator, one work, one pair. Whatever Futabatei's compliance turns out to be, it is one execution of one rule by one person.
2. The Japanese is a one-witness modernised text (workshop/canon/aibiki/manifest.md §Limitations) and the punctuation may have been normalised by the 1969 anthology. This is the largest single threat to condition B and no amount of alignment repair touches it. If Futabatei complies, the possibility that a twentieth-century editor made him comply is unrefuted here.
3. Neither Russian witness is the printing Futabatei used.
4. The lead's Japanese is modern and non-native. §7.
5. The permutation null shuffles counts across blocks within this work, so it holds the work's own distribution fixed. It is a null about matching, not about Japanese punctuation in general.
6. Nothing here bears on Tier D, which is NOT PASSED and is untouched. No quality judgement about any translation is made or elicited by anyone.

12. Pre-flight budget

One critic call and one three-seat alignment check. Worst case built from max_tokens (note (abc)): critic 8,000 at P4/qwen rates ≈ $0.13; three alignment seats at 3,000 each, priced at the worst plausible provider (S022) ≈ $0.09; one reserve ≈ $0.03. Declared worst case $0.25. Day headroom at session start: $4.592572223. The translation limb, the aligner, the analysis and the verifier are lead work and cost nothing (charter §3, A4).