Repository path: workshop/experiments/E-20260814b-honorific-hands/design.md · rendered 2026-09-09

E-20260814b — does a published human hand carry a Japanese honorific, and by what device?

Frozen 2026‑08‑14, ARM-honorific-hands step 1 (T5). The translation limb (T-genji-yomogiu-R06-v1 at f2cf142, T-genji-yomogiu-R04-v2 at ffa3a5f) was frozen before this file existed, in that order and in separate commits. No published English of this span had been opened when this design was frozen.

1. Question

At sites where 『源氏物語』 grades social footing in grammar English does not have, do published human translators carry the grading — and do they use the rank-noun-in-the-pronoun-slot device that RS-20260813e §4 saw only in 2026 machine hands?

framework/v0.2 §10.3 records the device and refuses to recommend it because "the only hands that do it are the two 2026 language models." §10.5 says the fix is materials: two or more freely readable published human translations. This design is that census.

2. Materials

Declared priming. During materials search on 2026‑08‑14, before the span was chosen, the lead read SUEM's rendering of §§3‑1–3‑3 — a different, non-overlapping stretch of the same chapter. This is why site selection is mechanical (§3): the lead must not be the one deciding which sites count. It is declared again on the translation artifact.

3. Sites — mechanical, 65, frozen with this file

extract_sites.py (frozen alongside) scans the source for a committed pattern list and emits every match with its paragraph, its speech/narration context and its stratum. The paragraph-level speech/narration state machine is part of the frozen rule. No site was added, removed or re-labelled by hand. Output: sites.tsv.

3a. Negative controls — mechanical, 10, frozen with this file

The 10 paragraphs of the span that contain zero pattern matches, computed by the same script and stored as controls.json: ¶4, 11, 12, 17, 18, 22, 23, 25, 26, 27. Two of them are the aunt's brusque commands to a servant (「さらば、侍従をだに」, 「いづら。暗うなりぬ」), two are narration that pointedly withholds elevation from the aunt (されど、行く道に心をやりて、いと心地よげなり), four are the two poems, two are speech tags. At a control paragraph the Japanese marks no social relation grammatically, so an English that marks one there has inserted it. Coders are given these ten interleaved with the 67 sites and are not told which are which.

4. Coding

Per site × hand, one of three codes:

• CARRIED — the English at the corresponding place marks the same social relation the Japanese marks at that site, by any means: a title or address noun, a deferential periphrasis, a register shift, or an explicit statement of rank.
• NOT-CARRIED — the content of the clause is rendered and the relation is not marked.
• NO-RENDERING — the clause the site sits in is not rendered by that hand at all.

NO-RENDERING is reported separately and excluded from the carriage denominator. This is method note (bnj) generalised from S180: a hand that omits a sentence has not lost the marking, it has omitted the sentence, and pooling the two would print an omitting hand as a losing hand. Note (bnj) was written for a locus absent from the translator's own source text; the same reasoning covers a locus the translator's source contains and his English does not render, and the two counts stay apart.

Seats and blinding. Two coders per hand, cross-assigned so that no model codes its own translation (RS-20260813e's rule): P2 google/gemini-3.6-flash on every hand; the second coder is P1c openai/gpt-5.6-terra except on P1's own hand, where it is P3c x-ai/grok-4.5. Coders see the Japanese paragraph, the marked form, a gloss of what it marks, and the hand's English — never the hand's identity, never the stratum labels, never the hypothesis, never any other hand's English. P5 deepseek/deepseek-v4-pro is out as a coding seat on any task shape, note (bne).

Consensus = both coders say CARRIED. Raw agreement is reported per stratum.

5. The device census — coder-free, and the sharper half

A committed regex over each hand's full English of the span, for the rank-noun set:

lordship · ladyship · my lord · my lady · your excellency · your highness · your majesty
your honour/honor · your reverence · your grace · his/her grace · liege · sir · madam · ma'am

Amended (critic BLOCKING 3). "Classified by position" was not a rule, and G3 was defined on the bucket that needed the judgment. Three fully deterministic measures replace it, all regex-only, all computable by device_census.py (frozen alongside), rates per 100 words of that hand's English of the span:

DEV-all − DEV-voc is reported as the non-vocative residue. No model is asked anything for this table, and no hand-classification enters any registered number.

6. Registered predictions

Bars set now, before any published English of this span is opened.

G3 is the prediction the arm exists to test and it is written as the lead's expectation so that its failure is legible. The decision rule is registered with it:

• G3 met (machines exceed humans by ≥ 1.0) → the confound is confirmed; framework/v0.2 §10.3's sentence stands and gains the measurement, and the device may not be recommended.
• G3 fails and the two rates are within 1.0 → the confound is dissolved; §10.3 may be restated as a description of English practice, still short of a recommendation because Tier D is not passed.
• G3 fails in the other direction (humans exceed machines by ≥ 1.0) → §10.3's stated basis is false as printed and the sentence is corrected, not merely qualified.

Amended (critic NON-BLOCKING 7). The dissolved branch is reworded so it does not overclaim what two pre-1930 hands can establish: "§10.3's 'only the two 2026 language models do it' is false for this span; period and hand remain confounded (§11), so nothing here says this is what English does." It can falsify the sentence as printed. It cannot establish English practice.

6a. Amendments made after the pre-run critic pass, before any hand or coder was dispatched

P4 moonshotai/kimi-k3 returned NEEDS-AMENDMENT, 8 findings, 4 BLOCKING. All eight are accepted; four in full, one in part.

7. Failure criteria — when a number may not be read

F1 is a live risk and is registered as such: SUEM is an abridging hand — his whole ch. XV runs about 12,850 characters of English against 14,226 characters of Japanese, which for a character-dense source implies heavy compression. If he compresses this span the way the count suggests, F1 fires on him and the primary rests on WALEY alone, which would be one published human hand and not two, and the result page must say so in its headline rather than in its limits.

8. Procedure

1. Freeze this file and extract_sites.py and sites.tsv. (done — commit recorded on the result page)
2. Independent pre-run critic pass: P4 moonshotai/kimi-k3, shown this design and the site table, asked for blocking findings. Amend in writing before any other call is dispatched.
3. Dispatch the two machine hands (P1, P3) on the source span, neutral prompt (§9).
4. Extract SUEM and WALEY's renderings of the span from the stored Gutenberg texts.
5. Run the device census (§5) — mechanical, no calls.
6. Dispatch the coders (§4), one call per (hand × coder), 65 sites per call.
7. Post-run verification: recompute every reported number from the raw outputs; mutation tests on the tally code.
8. Write RS-20260814b-honorific-hands.

9. The machine-hand prompt (frozen verbatim)

You are translating a passage of classical Japanese literature into modern English prose.

Below is a passage from 『源氏物語』, chapter 15 「蓬生」, in Shibuya Eiichi's critical text. Translate it into English. Keep the paragraph divisions exactly as given, one English paragraph per Japanese paragraph, and render the two poems as poems on their own lines. Do not add notes, headings, commentary or explanations. Output the translation and nothing else.

[source paragraphs]

Nothing in the prompt mentions honorifics, deference, social rank, register, or this design's question. temperature default, max_tokens 3000.

10. Pre-flight cost estimate

Built from max_tokens, not from expected output — note (abc).

Declared: $0.60. Today's UTC day stands at $0.000000000 of $5.00 before this run.

Declaration revised to $1.60 after the critic pass, and the reason is a seat failure, not a design change. P4 moonshotai/kimi-k3 was dispatched three times on the critic prompt. The first two spent their entire completion budget on hidden reasoning — 4,000 and 14,000 tokens — and returned finish_reason: length with empty content: $0.079782 + $0.233043 = $0.312825 for no output at all. The third, with reasoning: {"effort": "low"} and max_tokens 6,000, returned a full verdict in 91 seconds for $0.083293. The lesson is a rule of future conduct and is written as a method note: on a reasoning seat, max_tokens is not a ceiling on the answer, it is a ceiling on the answer plus the thinking, and the thinking will take all of it. Note (abc) says to build the worst case from the cap the request permits; it does not say that the cap can be consumed without producing anything, and this run is the first time that has happened here.

11. What this design cannot do

• One work, one chapter, one language pair. Two published human hands is the minimum that can separate a hand from a period, and it is still two.
• The two human hands are 44 years apart (1882, 1926) and both predate 1930, so a period norm of English translating practice is not separated from anything else. RS-20260813g names the same limitation on its own four hands and it is not solved here.
• SUEM translated into a second language and WALEY into his first. If they differ, this design cannot say which of the two differences explains it.
• Nothing about quality. Tier D is NOT PASSED. No verdict here says any hand's English is better than any other's.
• The lead's hand is not evidence. It is hypothesis-aware, written after the account existed, by the agent that designed the census.