Repository path: workshop/experiments/E-20260816e-mimetic-carriage/design.md · rendered 2026-09-09

E-20260816e — does anything in English carry a Japanese mimetic, and does it depend on whether the mimetic imitates a sound?

Frozen v1, 2026-08-16, before dispatch. Amendments after the pre-run critic are recorded in critic-response.md and dated; nothing below is edited silently.

1. Question

Japanese has a large open class of depictive adverbs — 擬音語・擬態語 — that English has no grammatical category for. RS-20260815e §4 found, on one story, one blind auditor and one call, that a single English sound-symbolic verb was credited with carrying a reduplicated mimetic at one site of six, against the R29 translator's log claiming six of six and "they cost nothing". That is a craft claim about a real translation problem and it is unreplicated.

This experiment asks a sharper question than "does it replicate":

When a translator marks a mimetic site rather than flattening it, is the marking credited with carrying the Japanese — and does that depend on whether the Japanese word imitates a SOUND or depicts a MANNER or STATE?

The conjecture is that English's resources are asymmetric: it has sound-symbolic words (hoot, rumble, splash, clatter, munch, snap) and it has almost nothing for depicting manner and state except phonaesthesia and plain repetition. If so, marking should be credited at sound sites and not at manner sites, and RS-20260815e's flat refusal would be the average of two different populations.

Why this is not method work (continue-prompt.md §4.5). One sentence: it measures what, if anything, an English translator can do at a site where the source uses a device English lacks, and what that move costs the English. Both halves are about translating literature. The site set is a grammatical class, which is the whole reason this device was chosen — see §2.

2. Why this device and not another: the site set is determinate by morphology

Two handbook instructions died in the two days before this session, both at the same place. §7.15's sound-figure instruction died because the lead's list of the Arabic's sound figures and three expert readers' lists agreed on 13 of 21 (RS-20260816c). §10.7's footing instruction died because two independent readers agreed at 0.471 on whether a line's footing is determinable at all (RS-20260816d). In both cases the instruction quantified over "the places where the source does X" and X could not be fixed.

A Japanese mimetic is a closed morphological form. A reduplicated bimoraic kana string (ABAB), or a bimoraic kana string with 〜り or 〜と, used adverbially or as the stem of a light verb. It can be found by rule. That is the property both dead instructions lacked, and it makes this the right device on which to ask whether an instruction is writable when the determinacy problem is removed. G2 tests that the set really is determinate rather than assuming it — note (bpt), which fired at S198 and is honoured here as a registered gate rather than a post-hoc diagnosis.

3. Materials

Source. 夏目漱石「坊っちゃん」(1906) chapter 2 whole, 5,981 non-space characters, Aozora 000148/752_14964. Public domain.

The base rendering. T-botchan-ch2-R06-v1 — the chapter rendered whole by the lead under R06 (lead single pass, source only), 3,407 English words, frozen at 8e1fbb24 with its translator's log and its fifteen-site census before this design existed and before Morri's chapter 2 was opened. Contamination against Morri 1918, measured after the freeze on the whole chapter: 19 shared 7-grams, 0 twelve-grams, longest common run 11 tokens — clean by tools/dependence_check.py's own verdict rule.

The two arms, built as minimal edits from the frozen base (materials/sites.json):

• MK — at each site the strongest available English depictive rendering: a sound-symbolic or phonaesthetic word, or a reduplicative/iterative construction. It must assert what the Japanese asserts and nothing more.
• PL — an unmarked lexical rendering asserting the same thing, with no sound-symbolic, phonaesthetic, iterative or coined element.

Outside the site the two strings are byte-identical by construction: each site is one en_frame with a single {X} slot.

MK-FAILED sites, declared before any call. Two of fifteen — M02 (ぼんやり) and M15 (ぐっすり) — carry the same string in both arms, because no depictive English rendering could be built without changing what is asserted. The reasons are written in materials/sites.json. These two are excluded from every carriage statistic. The FAILED count and its class distribution are a result in their own right and need no panel call — the same clause ARM-fluent-carriage registered at S189, and it is registered here for the same reason: if the marked arm cannot be built at manner sites, that is the finding, not a failure of the run.

Decoys (5). Sites in the same chapter whose Japanese contains no mimetic at all, given the same MK/PL treatment. D2 is the sharpest: an English sound-word (came clattering to a stop) at a site where the Japanese says only 威勢よく, "with spirit".

Parity plants (4). MK/PL pairs in which the PL side has been given a content error — a reversed direction, a negation, an opposite, a contradicted posture.

4. Seats and the disjointness this run can afford

config/models.md; slugs logged as provenance. P4 and P5 are out (notes (bne), (bps)), leaving five usable seats, and this design spends the fifth and sixth on keeping the raters disjoint from the panel — the separation S198 had to half-undo mid-run.

The census raters share no model with the carriage panel, which is the point: the class used in the interaction statistic is not assigned by anyone who judges carriage, and not by the lead.

Declared overlap, not hidden: P1 is both the pre-run critic and a carriage seat, and P2 is both the parity control and a carriage seat. Four usable seats do not permit total disjointness. The critic sees no bodies, and the parity task asks a propositional question with no carriage wording in it; both are recorded as limits and the primaries are recomputed seat-by-seat so a seat-specific verdict is visible.

5. Procedure — gates first, in this order

Note (boa): a control that can void the run is bought before the calls it would void.

1. G4 parity (19 calls, P2): each MK/PL pair, with the Japanese sentence, judged SAME or DIFFERENT in what it asserts. 15 real pairs plus the 4 plants, interleaved and unlabelled.
2. G2 census (2 calls, QR GL): the whole Japanese chapter, asked to list every 擬音語・擬態語, with the exclusion rule stated (degree/frequency adverbs and interjections out).
3. G3 class (30 calls, QR GL x 15 sites): each site classified SOUND / MANNER / UNSURE, from the Japanese sentence alone, one site per call.
4. G1 decoy (15 calls, 3 seats × 5 decoys): the carriage question at sites with no mimetic.
5. Main (39 calls, 3 seats x the 13 buildable sites): the carriage question and the English-quality question.

Every call: temperature 0, one item, arms presented as A and B with the MK/PL assignment alternating by site index so no seat can learn a position rule, and the mapping stored in the run row. Serial dispatch, append-and-resume, every attempt billed to the ledger (note (boe)).

6. The registered statistics

The unit is the cell (site × seat). Exact tails by enumeration where a tail is quoted.

Class assignment for P2 is the raters' majority, not the lead's. Where QR and GL disagree the site is dropped from P2 and the drop is reported.

P1's bar of 18 of 21 has exact one-sided tail P = 0.00019 against a 1/2 null; per seat, 6 of 7 has P = 0.0625. P3's direction is registered before the run because a post-hoc direction on a two-sided quantity is not a prediction.

7. What each outcome licenses, written before the numbers exist

• P1 and P2 both fire → the sound/manner split is real: an instruction can be written for the sound half and must not be written for the other. This is the arm's target result.
• P1 fires, P2 fails → marking is credited generally; the asymmetry conjecture is wrong and RS-20260815e's 1-of-6 does not generalise to this material.
• P1 fails → RS-20260815e replicates on a second Japanese source, a second story, three seats instead of one auditor, and the craft claim "English sound-symbolic words carry a mimetic" is refuted twice. This is a first-class result and the arm's completion criterion is met by it.
• G1 fails → nothing about carriage is said, and the finding is about the instrument, which goes to §Limits and not to the backlog (subject rule).
• Any outcome → the MK-FAILED count (2 of 15, both manner) and the Morri coding (§8) stand, because neither needs a panel call.

8. The free descriptive limb — Morri 1918 at the same fifteen sites

Yasotarō Morri's Botchan (Master Darling) (1918) chapter 2, public domain, opened after the freeze. His rendering at each of the 15 sites is coded by the lead into the same four categories (MARKED / PLAIN / OMITTED / WRONG-MANNER) and reported in the result page as an internal-judgment-only coding by one hand. It costs nothing and is not a registered prediction.

9. Failure criteria and things that void the run

• Any body unparseable after one mechanical re-dispatch → the cell is dropped and counted.
• More than 10% of cells dropped in any stage → that stage's statistic is withheld.
• G4 or G1 failing withholds what §6 says it withholds, with no renegotiation.
• Stop-loss reached → the run halts, whatever is bought is analysed, and the withheld statistics are named.

10. Budget

Worst case is built from the max_tokens caps the requests permit, not from expected length (note (abc)). Caps: critic 6,000; census 3,000; every item call 450, doubled on the one mechanical re-dispatch. The first cut of this table used a 1,200-token item cap and came out at a worst case of $1.27 — above its own stop-loss. It was re-costed until the inequality holds, which is what note (bpq) requires and what S198's critic caught there.

Worst case $0.83 < stop-loss $0.92 < declared ceiling $1.05 — note (bpq), two-sided. The UTC day opens for this session at $3.565384 of $5.00, $1.434616 unspent, so the ceiling fits with $0.38 to spare. The translation limb, the census, the contamination check and the Morri coding are lead work and cost $0 (charter §3, A4).

Note (bps) honoured: reasoning.max_tokens is probed on the longest item for QR and GL before anything else is dispatched, because the fix for one seat broke another at S198.