Repository path: workshop/experiments/E-20260817e-mimetic-reading/design.md · rendered 2026-09-09

E-20260817e — does the Japanese mimetic license the English phonaestheme? ARM-mimetic-carriage step 2

v3. Two independent adversarial pre-run passes, both NEEDS-REDESIGN, both answered in writing in critic-response.md; v1 and v2 were never dispatched. The design that runs is the third, and its central control was invented by the second critic's objection, not by the designer.

1. The question, and two designs that died before it

RS-20260816e §3: at a Japanese mimetic site an English rendering that enacts the sound or manner and one that states it are not paraphrases of each other. Eight pairs of eleven failed a propositional-parity gate. Every design that would price the marking against a constant was killed by that, and the arm registered the successor question before this chapter was translated:

If a depiction and a statement are different claims, then a translator at a mimetic site is choosing between two readings of the source. Which one do readers who can see the Japanese say the Japanese makes?

v1 asked it as a forced choice — which rendering asserts what the Japanese asserts, no more and no less — and the critic's BLOCKING 1 showed that this demands exact cross-linguistic identity, returns NEITHER almost everywhere, and makes the run vacuous.

v2 replaced the forced choice with two independent binary judgments per arm (ADDS, OMITS; SAYS = neither) and read the enacting arm's SAYS rate at real mimetic sites against its rate at eight chapter-3 sentences containing no mimetic word. The second critic killed that too, on two counts that are both right:

• BLOCKING 1 — the threshold sat inside the noise. At n = 15 and n = 8 the standard error of a difference of proportions is √(.25/15 + .25/8) ≈ 0.219, larger than the registered 0.20 bar. About 98 items per condition would be needed for 80% power. The bar could be cleared by sampling variation alone.
• BLOCKING 2 — "contains no mimetic word" is not "states no sound or manner." Seven of the eight null sentences state a manner outright — なるべく大きな声をして, 急いで, 威勢よく, 遠慮なく, 泳ぎ巡って, 追っ払う, 引き下がって. They were not a baseline; they were a second, uncontrolled set of items.

2. v3: the control the critic's objection implies

The comparison moves inside the item. The same two English arms, the same prompt, the same seats, are put to the Japanese sentence with its mimetic and to the same sentence with the mimetic deleted.

Three things follow, and they are the whole reason for the redesign.

1. Every item-level defect the critic named is now a constant. plunged in noisily is awkward, rattled and clattered is inflated, flimsy, flappy doubles the descriptive load — and each of them is identical in both conditions. They cannot produce a difference between the conditions. The critic's BLOCKING 8, its longest finding, is answered by construction rather than by argument.
2. Any demand characteristic in the wording is a constant too. BLOCKING 3 and 10 said the prompt invites the seat to find additions in the marked arm. It does — equally in both conditions. The leading example list was removed anyway, the English target span is now marked, and a reason is required per arm; but the design no longer depends on the prompt being neutral.
3. The statistic changes from an unpaired difference of proportions to a paired sign test, on which BLOCKING 1's power arithmetic does not bite in the same way: nine paired items flipping in one direction is p = 2⁻⁹ ≈ 0.002 exact, and the design registers the sign test in advance.

The entry rule for the paired set is grammaticality of the deletion, applied before any call: delete the mimetic and nothing else, and keep the item only if the Japanese survives as Japanese. Nine of fifteen qualify. The six that do not are named with their reasons in items.json and here, because a silently shrunken denominator is the failure this project keeps writing notes about:

One declared edit inside a deletion: at M08 the degree adverb やに is removed with にやにや, because 「やに笑ってる」 leaves an intensifier stranded on a bare verb.

What it teaches about translating literature (the subject rule, continue-prompt.md §4.5): it tells a translator whether the English phonaestheme at a mimetic site is licensed by the mimetic — by testing whether readers' judgement of that same English changes when the mimetic is taken out of the Japanese. The site set is a grammatical class; the readers are not the translator and are not one of the project's instruments being audited.

3. Materials

Built by materials/build_materials.py into materials/items.json; every chapter-2 row is copied verbatim from E-20260816e/materials/sites.json and every chapter-3 row is asserted against the frozen files it was transcribed from. 27 reading items:

The translation limb. T-botchan-ch3-R06-v1 — 「坊っちゃん」chapter 3 whole, 5,821 Japanese characters → 3,410 English words under R06, frozen at 673f0bed with a nine-site mimetic census and, at every site, the two renderings the translator was choosing between, written before this design existed. Its four buildable sites are the N items; its five unbuildable ones join chapter 2's four in the buildability stage.

Contamination, measured after the freeze and before this design was written. Against Morri 1918 chapter 3: 10 shared 7-grams, 1 twelve-gram, 0 fifteen-grams, longest common run 12 tokens, tools/dependence_check.py verdict DEPENDENT?. The run is "i was taken for a locksmith when i went to see the", rendering 「錠前直しと間違えられた事がある。ケットを被って、鎌倉の大仏を見物した 時は」. No figure in this design depends on independence from Morri: he is not a comparator here, appears in no item, and is read by no seat. Against the lead's own chapter 2 (note (bhb)): 1 shared 7-gram, longest run 7, clean.

The A/B slot for every (item, seat) pair was fixed by random.Random(20260817) inside build_materials.py before any call was made and is stored in items.json. The analysis reads the slot back rather than assuming it. The two conditions of a pair get independent slots, so a seat cannot recognise a repeat by position.

4. The seats

P2 google/gemini-3.6-flash, P3 x-ai/grok-4.5, QR qwen/qwen3.7-max for the reading task; P2 and P3 for the census, the class and the buildability second hand; P1 openai/gpt-5.6-terra for the two critic passes and nothing else. P4 and P5 are out on any task shape (notes (bps), (bne)); GL is out on any long prompt, which is why the chapter-3 census does not use it.

These are three specified systems, not a reader population (v2 critic, NON-BLOCKING 14, taken). Every claim below is a claim about what these three seats said on a curated Botchan item set. The phrase "the source licenses" is not used anywhere in the result without that qualifier.

5. Predictions, registered before dispatch

1. The enacting arm's ADDS rate rises when the mimetic is deleted. This is the experiment. The lead predicts it rises at 6 or more of the 9 pairs.
2. The lead predicts the enacting arm is already ADDS at a majority of the 15 real sites — against the lead's own instinct while translating, which was that the phonaestheme was the faithful choice. If prediction 1 holds and 2 holds, the reading is that the mimetic licenses the phonaestheme partially: it makes the addition smaller, not zero.
3. The stating arm's OMITS rate falls when the mimetic is deleted — for the same reason, mirrored. A design in which only the enacting arm moves is a design measuring one arm.
4. The identity items produce identical flags within a seat.

(Prediction 4 of v2 — a sound/manner class effect — has been struck on the critic's NON-BLOCKING 16: the design says it cannot test it, so it is not a prediction of this run. The by-class counts are reported as counts and refused as evidence.)

6. Gates, and what each one withholds

Missing-data and tie rules, registered (v2 critic BLOCKING 6). Three seats per cell; the majority is 2 of 3 and cannot tie. If a body is dead after one mechanical re-dispatch, that cell has two seats; if they split 1–1 the cell is unresolved, the pair is dropped under G3, and both the unresolved cell and the dropped pair are named. Unresolved cells never enter a denominator. G1 and G2 are read before the primary is computed (note (boa)).

7. Procedure

Serial, temperature 0, one item per call, max_tokens 600 against a reasoning cap of 120 (note (bpv)), 3,000 against 400 on the two census calls. run.py enforces the stage order census → classify → reading → build and refuses a stage whose dependencies are not complete; it writes a dead row after one mechanical re-dispatch so a permanently unparsable job is never re-bought; it checks the stop-loss before each call and the hard ceiling after each; and the reading parser requires all six lines, rejects a duplicated flag, and requires a reason for each arm. (All six are the v2 critic's BLOCKING 12, taken.)

Judgment is not parallelised. Every call is one seat on one item.

8. The primary

Primary — the paired flip. For each of the 9 pairs, the majority ADDS flag on the enacting arm in the mimetic-present condition and in the mimetic-deleted condition. Count:

• flips to ADDS (present NO → deleted YES) — the mimetic was licensing the phonaestheme;
• flips away (present YES → deleted NO) — the opposite;
• concordant pairs, in either direction.

Registered test: an exact two-sided sign test on the discordant pairs. With 9 pairs, 9–0 gives p = 0.004, 8–1 gives p = 0.039, 7–2 gives p = 0.180. The bar is not a rate threshold and there is no 0.20 anywhere in this design.

The same statistic is computed for the stating arm's OMITS flag, which is prediction 3.

Reported alongside, descriptively and with no test: the 4 × 2 table of ADDS and OMITS majorities for both arms at all 15 real sites; the per-seat marginals for every flag (the v2 critic's BLOCKING 5 — a raw agreement figure without marginals is uninterpretable); flag-specific pairwise agreement; and every WHY clause verbatim at the sites that flip, because on this project's record the reasons have been worth more than the counts.

What this design still cannot do, written before the numbers exist.

1. 9 pairs is 9 pairs. A 6–3 split is p = 0.51 and means nothing. The design can return a clear answer or no answer, and the second is reported as no answer.
2. It cannot separate "the Japanese says this" from "these three systems read Japanese this way." Charter §4 forbids treating panel agreement as validation.
3. It cannot test the sound/manner conjecture. 6 clean sound sites against 2 clean manner sites is not a contrast.
4. A deletion is not a natural sentence. 「汽船がとまると」 is grammatical Japanese but it is Sōseki with a word taken out, and a seat may be responding to the mutilation rather than to the absence of the property. Nothing in this design excludes that, and the result page says so.
5. M12 and M13 share a Japanese sentence. M13 has no clean deletion and so is absent from the primary; the clustering is resolved by that accident and not by design.
6. One work, one author, one language pair, one translator. Both chapters are Botchan.

Secondary, rebuilt blind (v2 critic BLOCKING 11). The nine sites where the translator could build no enacting arm go to P2 and P3, which are shown the Japanese, the span and the stating English with its slot marked, and are asked for up to three alternative renderings at least as accurate. The prompt names no lexical class, no theory, and no example. Every candidate is printed verbatim. Registered: a NONE answer has no evidentiary force whatever — a seat declining to better a phrase is not evidence that English lacks a resource — and any classification of a returned candidate as phonaesthemic is a lead judgment, marked internal-judgment-only where it appears.

9. Failure criteria, stated as failures

• G1 or G2 withholds → the primary is not printed, and the arm closes on that.
• Sign test p > 0.05 on both arms → no answer, reported as no answer, and the arm closes retired with the registered statement that the choice at a mimetic site is not shown to be recoverable from the source by this design at this n — not the stronger claim that it is a style choice, which the v2 critic's BLOCKING 7 correctly refused.
• Fewer than 7 pairs survive G3 → the sign test is not run and the counts are reported bare.
• Any stage crossing the stop-loss → the run halts and the result page reports what was not bought.

10. Budget

Pre-flight from the max_tokens cap each request permits (note (abc)), including the one mechanical re-dispatch at double cap.

Worst case $0.74 < stop-loss $0.82 < declared ceiling $0.95 (note (bpq), two-sided). The UTC day 2026-08-17 opens at $0.00 of $5.00 and this is its first session. The two critic calls cost $0.115 and killed two designs, which is what the money is for.

11. Decision log — investigator choices, not mechanical ones

The v2 critic's NON-BLOCKING 15 asked for this list, and it is right that the word frozen prevents later change without making a choice neutral. Free choices made by the lead: which spans count as mimetic (the census rule); which of the fifteen deletions are grammatical; removing やに with にやにや at M08; the three identity items; G2's two-of-three cutoff; the seed 20260817 and the slot schedule; which nine sites go to the buildability stage; keeping chapter-2 items the critic called defective rather than editing a frozen artifact. Inherited constraints, not chosen here: the eleven chapter-2 pairs and their exact strings; the chapter-2 classes; the panel composition; the budget cap.

12. Verification

analyse.py --json produces every number in the result page. verify.py recomputes them from run.jsonl by a route that re-parses the raw bodies rather than trusting the stored parses, does not import analyse.py, re-derives the slot decoding from items.json independently, and recomputes the sign-test p-value by exhaustive enumeration rather than from a table. Mutation tests: flip one reading flag, break one identity item's symmetry, delete one census word, flip one class, and one negative control that corrupts an unused field and must change nothing.