Repository path: workshop/experiments/E-20260813e-slot-typology-ja/design.md · rendered 2026-09-09

E-20260813e — where a source's footing mark lands in English, and whether position alone decides

ARM-footing step 2 (T5), the arm's closing step. Frozen before any English of 「煙管」 sections 三–八 was written, before any hand was dispatched, and before Glenn Shaw's 1930 English of this story was opened by the lead.

1. The question, and why it is not the question step 1 asked

RS-20260812i §3.1 offered, post hoc, an account of why Russian's deference clitic -с reaches English at 42 of 43 while the ты/вы choice reaches it at 0 of 11. Both are bound morphology; neither has an English equivalent. The proposed difference was positional:

The clitic sits at the end of an utterance, and English's utterance-final position is optional and open: sir drops straight into it. The ты/вы choice sits in the subject pronoun, and English's subject pronoun is compulsory and has exactly one form.

That account was written after the numbers were seen, and its own result page says so (§6.2). It has never predicted anything in advance. This experiment makes it predict, prospectively, in a third language family — and it is built around the one cell that can tell the account apart from its obvious competitor.

• Account A — positional freedom. A source mark reaches English if the English counterpart of the position it sits in is free: a device can be added there without changing what the sentence says.
• Account B — positional freedom and a conventional device. A free English position is not enough. English must also have a device that conventionally lives at that position. English has one at the utterance-final vocative (sir, my lord); it has none in the pre-nominal modifier slot, where your honourable pipe reads as parody rather than as deference.

The two accounts agree everywhere Russian and Bengali have been measured. They disagree on exactly one thing this project can reach: the Japanese honorific prefix 御/お. It attaches at the left edge of a noun; the English counterpart position — a pre-nominal modifier — is free by Account A's test, since the pipe → your gracious pipe adds no propositional content. Account A therefore predicts it is carried. Account B predicts it is lost, because no conventional English device sits there.

Whichever way that cell goes, framework/v0.2 §10 has something to say that is not a restatement of what was already believed.

2. Materials

Why this story. Selected on measured source-side signal before anything was designed: of six freely-available Akutagawa stories that Shaw also translated, 「煙管」 has the highest density of graded dialogue per source character. It also carries the single best site this project has met — the same word 手前 used by one speaker as a contemptuous second person to an equal and as a humble first person to a daimyō (C1-06/C1-11/C1-12 against C1-08/C1-13).

3. The four channels

Cut by where the mark sits in the source and by what English has at the counterpart position.

C4 is a positive control and is declared as one in advance, not a finding. A lexical honorific is also a thing that was said; predicting that it survives is predicting that translators translate words. This is step 1's pre-run critic's finding against its own P1, imported here rather than re-learned.

4. The census, and how sites were adjudicated

census.py emits candidates from patterns committed in this file. adjudicate.py applies an explicit include/drop table to those candidates and writes sites.json. Both were run, and the site list frozen, before any English was written, opened or dispatched.

Adjudication rules, in order:

1. A candidate is a site only if its form grades the speaker against the addressee — up (deference, self-lowering) or down (contempt, condescension).
2. C2 counts one site per speech, classified by the level of that speech's final predicate, and only where the level is marked. The plain forms peers use with each other (だ さ よ ぜ) are the default and are not sites; ございます-class forms (up) and the lord's じゃ/ぞよ (down) are.
3. C3 admits only true honorific prefixes. お in お前, おい, おお is not a prefix and is dropped.
4. C4 admits only marks whose removal changes what the sentence says. 居りまする (the humble auxiliary おる) is contentless suppletion and is dropped from C4 rather than inflating a control.
5. Direction (up / down) is recorded per site from the speaker–addressee table in sites.json.

Frozen counts: C1 13 · C2 12 · C3 7 · C4 19 — 51 sites in 44 speeches.

5. Hands

LEADb/LEADa is the internal contrast this design gets for free: one translator, one story, one half written before the hypothesis existed and one after.

6. Coding

Carriage is coded by two panel seats that produced no hand: P2 google/gemini-3.6-flash and P5 deepseek/deepseek-v4-pro. Plain-text line output, not structured output (NEXT.md's standing finding on P5). Each coder sees, per hand: the source utterance, the marked form, a literal gloss, and the hand's English for that utterance — and never the hand's identity, never the channel labels, and never any statement of the hypothesis.

The verdict set, fixed here:

• CARRIED — the English contains a device whose removal would leave the sentence grammatical and propositionally unchanged, and whose presence marks the same direction of footing as the source's mark.
• LOST — no such device.
• REV — a device marking the opposite direction.
• NA — the hand does not render this utterance at all.

The lead codes all sites as well. The lead's coding is a third reading and enters no primary.

7. Registered predictions

Pooled over hypothesis-blind hands (SHAW, LEADb, P1, P3), NA excluded from denominators.

G3 and G3′ are the two accounts. Both are registered. A C3 carriage between 0.15 and 0.50 decides neither and will be reported as deciding neither.

8. Failure criteria

9. Procedure

1. Freeze this design, census.py, adjudicate.py, sites.json. Done before step 2.
2. Independent pre-run critic — an off-panel seat (qwen/qwen3.7-max) that produces no hand and does no coding. Amendments recorded in amendments.md before any other call.
3. The lead translates 三–八 from the Japanese alone; translator's log frozen with it.
4. Dispatch P1, P3 on the whole story. Prompt names no channel and no hypothesis.
5. Overlap checks (F2, F6) by tools/dependence_check.py.
6. Dispatch coders P2, P5, five hands each.
7. Compute, verify by an independent recomputation with mutation tests, write the result.

10. Pre-flight cost

Built from the caps actually sent, never from an expected output length — note (abc).

Today's headroom at freeze: $5.00 − $3.450609849 = $1.549390151. The ceiling fits.

11. What this cannot show

Tier D is not passed. Nothing here is a claim that carrying a mark is better than losing it, or that Shaw's English is good. The question is entirely where the marks land, and the framework section it feeds says so in its first line.