Repository path: framework/v0.2/README.md · rendered 2026-09-09

Framework v0.2

FROZEN AS THE RECORD at v0.2.52 (2026-09-04, S244, on Tom's direction — PROJECT.md §11). No new numbered section is added to this page. Every finding after S243 goes into an entry of framework/v0.3/, the handbook organised by translation problem, which consolidates this page family by family (framework/v0.3/README.md §Index maps every §7.n below to its entry). Nothing here is deleted; sessions grep the section an entry consumes and never read this page whole (380 KB).

The second release, 2026-08-08 (S133), ARM-marking-work step 2. It changes one thing and it is a subtraction: prediction 1 is retired after five attempts in five language pairs, and what replaces it is a statement about how rare the problem R1 addresses turns out to be.

v0.1 is not superseded and is not withdrawn. Its recommendation R1 stands with its text unchanged; its §5 (the fourteen candidates), §6 (coverage without a scalar) and the whole of its evidence base carry over unaltered and are not restated here. Read framework/v0.1/README.md first; this page records only what v0.2 changes and why.

Standing. provisional: true — Tier D is NOT PASSED, so no recommendation here may depend on anyone's judgment of whether a translation is good (v0.1 §1; evidence class X3 remains inadmissible). This is an internal versioned artifact. It is not a publication: the repository is private and external release is Tom's alone (charter §7, §10).

1. What changes

Nothing else in v0.1 changes. No recommendation is added or withdrawn.

2. Prediction 1 is retired

v0.1 §3 prediction 1. On fresh Class A sites in a new pair, R1 recovers a marking at more than half.

Status in v0.2: RETIRED. Not falsified, not discharged — retired, because five attempts have failed to produce the population it quantifies over, and the fifth showed that the admission condition proposed for a sixth is not identifiable.

The reason it is retired rather than re-worded. ARM-marking-work was constituted to write a 1′ whose admission condition was defined on the source's marking and the target's device rather than on a reader's recovery — the repair RS-20260806e §7 named. S128 supplied the candidate: source marks + content discordant + target device available. S133 tested the middle term and it did not survive. RS-20260808b §4.2:

A per-utterance admission condition cannot pick out a phenomenon that lives in a text's trajectory. «Я, ваше превосходительство… Очень приятно-с!» is called concordant by 3 of 3 readers of the Russian, and they are right: on its own, that sentence is a man deferring. It is discordant only if you know that the same man said «Миша! Друг детства!» ninety seconds earlier.

S128's κ = 0.786 was earned on an Ōgai courtroom scene in which the standing is restated in every line — a text where the situation is inside the sentence. That is a property of that text, not of the phenomenon.

Retiring a prediction is not retiring the recommendation. R1 tells a translator that the option exists; it has never claimed the option is often needed, and v0.1 §2 is explicit that it cannot tell anyone to take it while X3 is inadmissible. What retires is a quantitative claim the project kept failing to set up.

S1 — what replaces it, stated as a finding rather than a prediction

S1. Across five language pairs, this project has not produced a population of sites at which a competent English rendering loses a relation the source marks grammatically. Where the source's marking rides on a device English also has — a noun phrase of address, a courtesy formula, a title — the relation transfers whether or not the same-category device is carried. On the evidence to date, the loss R1 exists to repair is uncommon, and a translator applying R1 should expect most sites to need nothing.

Evidence, in class X1a and X2, no jury:

• DE→EN (RS-20260806e). King 1914 carries no English device of address at 9 of 9 vocative-free asymmetric-address sites in «Michael Kohlhaas»; his device-free English is judged to convey the relation at 8 of 9. Oxenford 1844 carries a device at 4 of 9 and also conveys at 8 of 9. Published practice splits on the device and not on the relation.
• RU→EN (RS-20260808b §4.1). At the seven utterances of two whole Chekhov stories that three readers of the Russian put at the floor of the standing scale, Constance Garnett's published English — which deletes the deference clitic at every one of them — transfers the standing to blind readers of the English alone at a mean error of 0.048 scale points out of seven. Two independently generated arms match her to three decimal places.
• JA→EN (RS-20260807d §4.2). A close literary rendering carried no English device at 33 of 51 marked sites, and the phenomenon that would make that matter occurs at ~8% of them.
• PL→EN (RS-20260805c). The registered admission gate admitted 0 of 8: the filed close translation already conveyed the relation at every site.

What S1 does not say. It does not say the marking is unimportant, and it does not say nothing is lost — what it says is that nothing is lost that this project's instruments can detect at the level of an utterance. RS-20260808b §4.4 is the counter-case, n = 1: at «Что-с?», an utterance consisting of one pronoun and one deference clitic, the three arms that dropped the clitic all transferred the standing wrongly and the one that kept it (What, sir?) did not. Where the source has nothing but the marking, dropping the marking costs. That is one site, in one pair, from an arm excluded from every primary for contamination — a place to look, not a finding.

S1 gets no predictor, and ARM-carrier closes with the refusal written — 2026-08-12 (S166), RS-20260812d-slot-or-carrier

ARM-carrier was constituted at S161 to turn S1 from a description of five failures to find a loss into a statement with a predictor in it — or to fail trying and say so. It failed trying, twice, and this is the saying so. No predictor enters S1. S1 stands exactly as written above.

Step 1 (RS-20260811g, Turkish) was stopped by its source-parity gate: three arbiters could not agree that the two Turkish variants differed, and every primary was withheld. Step 2 (RS-20260812d, Japanese 「祝盃」) repaired all three of the things step 1's §8 asked for — a source-side panel admitted on a rule registered before dispatch, windows selected on measured source-side signal before any hand was paid, and fifteen windows with a graded response instead of six with a binary. Every one of those gates passed. The run was then stopped by the gate nobody thought was at risk.

G3, the floor — one hand translating the same passage twice at temperature 0.0 — came back at 0.8333 against a registered bar of 0.60. The two renderings are not the same text: across six pairs their similarity runs from 0.0229 to 0.8511. A post-hoc control, unregistered and licensing nothing, put the same arbiters on byte-identical English pairs and got 0 at 12 of 12, so none of the floor is the instrument and all of it is the hand. Per the design's own rule a failed instrument gate withholds every primary, and it does.

What the framework takes from this, and it is a constraint rather than a recommendation:

Before any design proposes to detect a shift of the size these questions turn on, it must measure what one translator's own re-rendering of the same source moves — and set its bar above that. On this run that quantity is 0.8333 on a 0–3 scale, a third of the way to "a clear difference", from a hand asked twice at temperature zero. Every arrival statistic this project has published on marked relations — RS-20260811f's 0.0000, RS-20260811g's 0.0556, this run's five cells — is a number whose floor was assumed rather than measured, and only two of those runs measured a floor at all.

Nothing about carriers or positions is written into §2, and the withheld pattern is on the result page where it can be checked but not cited. ARM-carrier closes retired at 2 of a declared 2, with the reason written, per its own Done when.

3. §4 gains a third device-category contrast, and it is inside a single utterance

v0.1 §4 carries two contrasts: pronoun against adjective (RS-20260805h), and pronoun against address noun (RS-20260806e, "alternatives, not ingredients"). The third is cleaner than either, because the two categories are simultaneous in the same words.

Russian marks Chervyakov's standing twice over in the same breath: with the noun phrase «ваше превосходительство» / «ваше —ство», and with the clitic «-с» glued to whatever word is handy — «я чихнул-с», «брызнул-с», «Хи-хи-с», «Что-с?».

English has the first category and does not have the second at all. Garnett carries every noun phrase (your Excellency) and drops every clitic. The measured result: transfer error 0.048 at those sites. The noun phrase carries the whole load and the clitic is redundant on top of it — until the noun phrase is absent, and then the clitic is the only thing there (§2's «Что-с?»).

What this adds to the "alternatives, not ingredients" finding. RS-20260806e established that applying two categories does not mark twice. This adds that the categories are not equal in reach: a device that can only attach — a clitic, a suffix, a particle — is invisible when a device that can predicate is available in the same clause, and is load-bearing when it is not. A translator deciding whether to compensate for a lost particle should first ask whether the utterance carries an address noun; if it does, the particle is very likely free to lose.

And this is the first time the two categories have been separated with the propositional content held exactly fixed, because they occur in the same sentence and one of them has no English counterpart to carry.

4. §7 — the RU→EN row, strengthened

v0.1 §7 records RU→EN as evidenced, machine-read, on RS-20260802e sites S2, S3, S4 — three sites of Turgenev. v0.2 replaces the row's evidence, not its standing:

Every other row of v0.1 §7 is unchanged, including EN→anything, untested, the standing gap since S015.

5. §8 Q-b — third and fourth measurements, and the question changes shape

Q-b as v0.1 states it: how much of an address system survives a good translation at all? — and, sharpened at S121–S122, where in a text does the marking do work that the scene does not already do?

Q-b's shape changes and the release says so. The question was posed as how much survives. Four measurements say the interesting quantity is the other one: how little there was to lose. In three of the four, a published or close rendering carried little or none of the source's device and conveyed the relation anyway. The one place the question still bites is where the source has only the device — §2's «Что-с?», and RS-20260806e's A2, where the speaker's content works against the deference his address claims. Both are n = 1, in different languages, five sessions apart, and they are the same kind of site.

Q-b stays open on that residue and on nothing else.

6. §8 Q-d — new: is the marking that matters a property of trajectory rather than of utterance?

Opened 2026-08-08 by RS-20260808b §4.2, and it is where the retired prediction's evidence actually points.

Every measurement this project has made of relational marking has been per site or per utterance — a sentence, a span, a quoted turn. Chekhov's «Толстый и тонкий» is built on something no per-utterance instrument can see: the same speaker addressing the same person with ты and then with вы, four paragraphs apart. Three independent readers, shown each utterance alone, correctly called both halves concordant, because each half is.

The question. Where a source marks a relation by a change in form across a text — a switch of pronoun, an onset of honorifics, a drop of a title — does English carry the change? And what does a published translator do with it?

Why it is answerable, and cheaply. The change is a textual fact, countable in both languages with no jury: Oxenford's 48 archaic pronouns in the Luther scene and zero in the Herse scene (v0.1 §4) is already a measurement of exactly this kind, made as an aside. RS-20260808b's corpus supplies a second: the thin man's switch happens at a nameable paragraph in Russian and Garnett's English has to do something there or nothing.

What Q-d must not repeat. It is a question about a text's arc, so it may not be answered by putting isolated utterances to readers. Whatever instrument it uses has to show the reader the trajectory — which is the opposite of the blinding that every design since S101 has needed, and that tension is the design problem, stated here so a later session does not discover it at the critic stage.

Q-d, ANSWERED IN ONE PAIR — 2026-08-09 (S140), RS-20260809a-trajectory-ja

Russian→English, status released-on-repair. Three independent hands do not put a source's pronoun-borne change of address into their English in any form a reader recovers — +0.134 against a null-favouring bar of ≤ +0.40 — while the same rater seats, on the same passages from the same hands, recover a vocative-borne change at +0.786.

The generalisation Q-d can now carry, on one pair: what predicts whether a trajectory-level relational change survives into English is whether the source's device has an English exponent, not how much of the source it occupies. A Russian pronoun switch has none and is lost; a vocative has one and is kept.

Three qualifications travel with it and are part of the claim.

1. released-on-repair is weaker than a blind release. The figure was measured at S139 and withheld when the source-side gate returned +0.667 against +1.00; S140 re-measured that gate at parity — byte-identical texts, the same prompt, the same bar, 6 ratings per site instead of 2 — and it returned +1.389. The release rule was registered before the re-gate was dispatched and the bar was never moved, but the gated quantity was already visible when it was written. A reader who thinks that decisive should treat +0.134 as still withheld.
2. The two-pair requirement is UNMET. Japanese→English was attempted at S140 on the densest device the language has — every finite predicate in the tail moved between です・ます and plain — and both its gates fired: the source-side gate returned +0.139 against +1.00 on twelve ratings per site from six seats, and the positive control returned −0.125 against +0.75. No primary is licensed. Read that as the instrument could not establish legibility, not as a fact about Japanese.
3. A second, gate-free observation now has two runs behind it. Readers of the English are asked what words moved them, and they overwhelmingly quote events, not address: 2 of 36 quotes on the Japanese control name the manipulated constituent, and 2 of 36 on the Japanese primary contain a contraction — while the mechanical census over the same English shows contractions moving +2.639 per 100 words in the manipulation's direction. The hands encode the change and the probe does not see it. That is a property of the reading probe, and any successor design owes it an answer before it spends again.

Q-d stays open on the second pair, and only on that.

7. §8 Q-e — new: register carriage cannot be recommended, and this is exactly why

Opened 2026-08-08 by RS-20260808f §5, closing ARM-low-pole. This section is the arm's answer to the question it was constituted on: can framework/v0.2 carry a recommendation of the form "where the source marks low register, do X"? It cannot, and the obstacle is now specific.

What is settled enough to state as a problem. Across four published hands and four language pairs, at 47 sites where the source drops below its own neutral written register, no hand goes below its source on average anywhere, and the only four sites where any hand read below are all four direct speech — not one narration site in any cell (RS-20260808e §7, descriptive and resting on no withheld quantity). At the eight Italian narration sites, four low arms were measured in one call: the two allowed located idiom or phonetic spelling went below the source (−0.4167, −0.3333); the two forbidden them did not (+0.3750, +0.1667), and one of each pair was written by a model that had never seen the hypothesis.

Why that is not yet a recommendation. The constraint that produced the effect bundles three devices — regional idiom, class-marked grammar, and eye-dialect — and this design cannot say which one costs the register. It matters for practice more than almost anything else the framework could say here: a translator can adopt phonetic spelling without relocating a book, but cannot adopt them that or a wrong 'un without putting a Sicilian fishing village in England. Dropping the two sites where the unconstrained arm used phonetic spelling takes the effect from +0.7917 to +0.2778, inside the design's own registered futility band.

The design that would settle it was named here and was run at S145 (RS-20260809g-device-cross). What it found is in §7.1, and it does not license the recommendation either — for a different and better reason.

And one methodological result Q-e must carry rather than rediscover. ARM-low-pole was blocked in both its sessions by a content-parity gate, under two different phrasings, the second being the repair of the first. The reason is measured: at register-marked sites the checker flags the published hand as omitting or misstating at 0.6809 and as supplying at 0.7660 — the highest supply rate of any arm in the run, above a 2026 model told to write low. A content-parity gate cannot license a register claim about register-marked sites, because at those sites nothing passes it, including the hands the claim is about. Note (bkr).

7.11 The elevation half of Q-e, measured at last — and a light register rule is loudest where it has least room — 2026-08-15 (S188), RS-20260815-register-room

ARM-elevation-resolution step 2, and the arm closes resolved at 2 of 2. §7.4 stated Q-e's open half in one sentence: "where published hands sit above plain English is still unmeasured." That sentence is still true, and this entry does not answer it — no published hand is an arm here. What is now measured is the thing §7.4's sentence presupposed and nobody had checked: what a declared register policy does to the English at all, and where.

The finding, on Chekhov's «Злоумышленник» (1885, Russian), three complete lead renderings, twelve mechanically selected sites, three seats blind to arm, 36 of 36 bodies usable:

A one-step-up register policy is read as raising the source at its low places — and it is loudest at the shortest ones. Against an unruled pass at −0.25, the light-rule arm sits at +0.50 where the Russian drops in a 26–60-word speech and at +0.75 where it drops in a one-to-six-word line, on a −1/0/+1 source-relative scale.

Both explanations this run was built to separate are refuted, and the registered numbers say so. RS-20260814d deposited "a one-step policy is not experienced as raising the source" (A); its own §6 offered "it had nowhere to go inside a one-word shout" (B). Here Q1 holds at +0.7500 against a +0.25 bar, which refutes (A); and Q2, which predicted the short-item null would reproduce, fails at +1.0000 — the effect is larger at one-word items — which refutes (B). The cutoff-free check agrees: Spearman between span length and effect is −0.7178.

What a translator can take from this, stated at the size of the evidence. A moderate register policy is not a wash. Its audible work is done by grammar, not vocabulary, and by discrete non-standard forms rather than by volume of rewriting:

1. One corrected pronoun case moved every seat a full point. «Мы, народ…» is Us, the people… with no rule and We, the people… with the rule — one token in fourteen — and three blind seats coded the difference at +1.000, naming the case error unprompted. On the same eight sites, a 60-word turn with a quarter of its tokens rewritten moved them +0.333. The rewrite-fraction account was tested post hoc and is not supported (ρ = +0.345).
2. Contractions are the carrier. The frozen translator's log registered 49 contractions in the peasant's turns under no rule and 0 under the light rule, and named that as the arm's main device before any evaluation existed; the seats cite contractions in their own words at two sites. The log and the blind readers agree, which is rarer in this project's record than it should be (RS-20260812d, RS-20260806d).
3. A light rule and a heavy one are different policies, not two doses. Full ennoblement is at +1.000 in 36 of 36 cells and its effect is not concentrated at the low places — +1.250 at long low sites against +1.333 on ordinary prose — because its rule set raises the neutral by design. The light rule leaves the neutral alone and is invisible there except where it touched a word: its whole ordinary-prose effect came from two single-word narration changes, one of them logged as a deviation from its own rule, and the seats saw both.

And a result about the project's own low-site censuses that belongs to translators too. Three seats reading this Russian alone were unanimous on 55 of 57 spans with zero no-majority spans; the same three on Andersen agreed at 0.37–0.59 and split three ways on 7 of 27. Where a source marks register in its vocabulary, "where does this text drop below itself" is an agreed question; where it marks register by tone, it is not. RS-20260808e's 47 low sites rest on a three-annotator majority, and the reliability of that procedure is now measured at both ends and is a property of the source, not of the panel.

Evidenced on: Russian→English, one story, one author, one translator, three model seats. internal-judgment-only, provisional; Tier D remains NOT PASSED; the word reader is not used of these seats. Nothing here is about published practice or about quality.

7.12 The unlicensed-mark diagnostic reaches SOUND, and the first thing it finds is that sound is the wrong channel to be watching — 2026-08-16 (S191), RS-20260816-answering-figure

⚠ ITEM 3 OF THIS SECTION IS WITHDRAWN — see §7.14 (2026-08-16, S196). Its heading is also wrong as written: sound is not the wrong channel to be watching. Item 1, item 2 and the measurements stand; the post-hoc structural explanation in item 3, and the instruction the sound you add on top is heard as yours, were tested and failed.

§7.7 and §7.8 built the diagnostic on punctuation and on emphasis: count the marks your reader sees, and look at how many the source licensed. Neither said anything about the case a translator actually agonises over — the source has a figure the target cannot reproduce, and the handbook says compensate: put a device of your own at that place.

RS-20260816 put the move to a test built for it. Ibn al-Muqaffaʿ's «باب السائح والصائغ» from «كليلة ودمنة» was rendered whole under a regime that refuses English sound-patterning everywhere (R30), its fourteen Arabic sound figures enumerated from the Arabic before any English existed; then two arms, differing from that base and from nothing else, added one English sound device at each of the 13 places the Arabic has a figure (R31) and at 13 matched places where it has none (R32), same device classes, same counts, substitutions only. Three blind seats, 372 bodies.

The registered primary was withheld by its own gates and the gates are the finding.

1. A regime that refuses sound can be executed, and this is the first time the project has demonstrated it. The base rendering was heard as unpatterned at 0 of 26 cells.

2. Ornament can be relocated; it cannot be invented at constant length. Of thirteen compensations, three seats heard 4; of thirteen inventions, they heard 0 (exact one-sided P = 0.0478). The four heard are all multi-member parallel structures — two matched cola, an adjacent rhyme at a clause junction, a three-member alliterating construct. Every invention is necessarily a two-word pair, because a plain sentence offers no matched members to hang a larger device on. Caveat that must travel with this: the decoy arm was forbidden to add words, and a real ornamentalist adds words (RS-20260814g measured Burton doing it). This says nothing about supplied ornament in the wild.

3. What moves a reader's belief about the original is the surviving parallel STRUCTURE, and this is untested and post hoc. Asked was the original doing something with sound here?, the seats fired at 7 of 13 compensated figure sites, 2 of 13 uncompensated ones — with no sound anywhere — and 0 of 13 plain sites. On third-party English, two published Victorian hands separate the Arabic's sound loci from its plain ones at 6 of 8 against 0 of 8 (exact one-sided P = 0.0035). Every affirmative judgement in the run is justified by the seats in terms of parallelism, not sound: "parallel matched phrases", "matched word-shapes", "rhythmic parallel". The cleanest cell is Burton at the locus where he drops the second member — the inference dies there (N N N) while Lane, who keeps it, carries it.

What v0.2 will and will not say.

• It will not say "compensate". On the only measurement the project has, a substituted device that preserves the sentence's meaning is inaudible at nine loci in thirteen, and where the source is plain it is inaudible at thirteen in thirteen.
• It will not say "do not compensate" either. RS-20260815 §6 shows a published hand doing it at 3 of 4 and being heard.
• The diagnostic it does carry, in §7.7's own form — count, and look at the difference — now extends to sound with one addition: count the source's matched MEMBERS, not its sounds, and ask how many survive in your English. RS-20260814 measured on a different Arabic work that the structure crosses and the sound does not; this run's candidate consequence is that the part that crosses is the part the reader reads as the original's.
• The instruction a practitioner can act on today, and it is a diagnostic, not a rule: where the source pairs, triples or repeats, the members are doing the work your reader will attribute to the author — keep the count, the adjacency and the matched shape. The sound you add on top is heard as yours.

What is untested here (D-20260724-04): item 3 entirely. It was assembled from free-text reasons after the run, its registered primary is withheld, and the asymmetry in item 2 is equally consistent with the lead's decoy devices simply being worse than its compensations — the run's own arm-equivalence gate failed, and RS-20260816 §10.3 says so in those words. Two successor designs are named on that page. Tier D NOT PASSED; the seats are models, not readers.

7.14 §7.12 item 3 is WITHDRAWN, and the warning that replaces it is sharper — 2026-08-16 (S196), RS-20260816b-invented-figure

§7.12 was written yesterday and its item 3 said, marked untested and post hoc, that what moves a reader's belief about the original is the surviving parallel STRUCTURE, and that "the sound you add on top is heard as yours." RS-20260816b gave that account its first real test and it fails.

The test: two arms of the same English translation, token-identical except for one word at eleven of thirteen loci — the same added members, the same word count (+52 and +52 over the thirteen), the same propositions — differing only in whether the ornamented words chime. 402 blind bodies, three seats, all four third-party controls unanimous for the third session running, base arm heard as unpatterned at 0 of 13.

1. The sound does the work, not the members. Asked was the original doing something with sound here?, the seats fired at 9 of 13 on the chiming arm and 4 of 13 on its de-sounded twin (Δ = +0.3846; exact enumeration over 2^13, score 0.0312). The base, with neither, fired at 0 of 13. The members do move it — 0 → 4 — but the sound moves it more than twice as far.

2. The refusals are explicit and they name the thing §7.12 credited. Both arms carry the same parallel structure. On the de-sounded arm the seats wrote "only ordinary parallel phrasing" and "English offers parallel phrasing, not conspicuous sound pattern" — they see the members and decline to infer the author from them. Change suit to befit, choice to fair, borne to endured, proof to test — one word, nearly the same sense — and three blind seats begin telling you the Arabic had a sound figure at a place where it has none.

3. AND THE READER CANNOT TELL A COMPENSATION FROM AN INVENTION. At the thirteen loci where the Arabic has a figure, a genuine compensation is read as evidence of the original at 8 of 13. At thirteen loci where the Arabic has nothing, invented ornament is read as evidence of the original at 9 of 13. The gap is −0.0769, inside the 0–2 locus spread this instrument shows between sittings. On these materials the inference carries no information about where the source's figures are.

4. §7.12 item 2 survives, with its second clause now explained. Ornament can be relocated; it cannot be invented at constant length — the constant-length invention arm scored 1 of 13 again this sitting. It could not be invented because there was no room for it, not because invention is inaudible. Given the words an ornamentalist actually takes, invention is heard exactly as compensation is.

What v0.2 now says, and it is a warning rather than a recommendation.

• §7.12's item 3 is withdrawn in full. Its practical instruction — keep the members; the sound you add is heard as yours — had the second half backwards. The sound you add is heard as the author's.
• It still will not say "compensate", and now it says why. A supplied sound device tells your reader that the original was doing something there. It tells them that whether or not it was. That is the §7.7 diagnostic — a mark the reader can see that the source's reader could not — arriving at sound with its sharpest case yet, because here the mark is not merely unlicensed: it is read as a licence.
• The diagnostic a practitioner can act on, unchanged in form from §7.7 and now evidenced on sound: count the sound devices your English contains and count the ones the source licensed. Where the second number is smaller, your reader is being told something about the author that you invented.
• Keeping the members is still not refuted as craft. It is refuted as an explanation of the reader's inference. §7.12's structural observation — the members cross for free where the sound does not (RS-20260814) — is untouched.

What is untested here (D-20260724-04): item 3's own successor, the Q1/Q2 threshold gap. The seats called the chiming arm conspicuously sound-patterned at only 3 of 13 while inferring the original's figure at 9 of 13: a device can be too slight to be called ornament and still be enough to make a reader believe the author was ornamenting. That is post hoc and unregistered. One work, one chapter, one hand, one language pair, n = 13, exact synonymy unavailable, Tier D NOT PASSED; the seats are models, not readers. Limits in RS-20260816b §10.

7.15 The compensation diagnostic survives, and its denominator does not: "the source's sound figures" is not a determinate set — 2026-08-16 (S197), RS-20260816c-checked-ornament

NARROWED 2026-08-16 (S202) by §7.17. On a fresh chapter, with the question put as an obligation on the translator rather than as a description of the Arabic, the readers' target set is determinate and the thing that was rule-dependent was exclusion (a) of the inventory rule below, which §7.17 strikes. The instruction say which rule fixes your target set stands; the claim that no determinate set exists does not. Read §7.15 then §7.17.

§7.14, written earlier the same day, gives a practitioner one thing to act on: count the sound devices your English contains and count the ones the source licensed; where the second number is smaller, your reader is being told something about the author that you invented. This section is about the second number.

Three seats were shown the Arabic alone, one stretch marked, and asked whether the Arabic was doing something with sound there — saj', jinās or muwāzana. Twenty-one loci of «باب الناسك والضيف», labelled by an inventory frozen before any English of that chapter existed. The registered gate was that the inventory's figured loci and the unfigured ones would separate by 0.40. They separate by 0.1886, and the run's source-visible arms are withheld by their own rule.

The seats are not failing to read Arabic. They agree with the inventory at its clearest cases — the saj' chain on ‑jad at 3 of 3, the root-echo kathīrah / kathrah · athmār / thimār at 3 of 3 — and they part company on three rules:

1. Repeating a word unchanged is not sound work to them. The chapter's key repetition, «تركت لسانك … على لسان العبرانية», scores 0 of 3; one seat writes "repeated لسان is lexical repetition, not sajʿ, jinās, or muwāzana." Repeating a root in a changed shape they hear at ceiling. The inventory puts both in one class.
2. Rhyme carried by a pronoun suffix is saj' to them and is excluded by the inventory by rule. Every affirmative they give at an unfigured locus cites one — ‑hā, ‑atihi — and at one such locus all three agree.
3. Two matched imperfect verbs are muwāzana to the inventory and nothing to them, 0 of 3 at both such loci — the same objection the pre-run critic raised and the design overruled, arrived at independently.

What v0.2 now says.

• §7.14's diagnostic stands. Its denominator must be declared. The ones the source licensed is a count under a rule, and two defensible rules — this project's inventory and the one three strong readers of the source apply — agree on only 5 of the inventory's 10 figures and disagree on 8 of 21 loci. A practitioner who runs the diagnostic without naming the rule has not measured anything stable.
• The recommendation this blocks is the positive one. Where the source has a figure you cannot reproduce, put a device of your own at that place presupposes that "the places" are given. On this chapter they are not, and a translator following one rule invents where the other says compensate. v0.2 still will not say "compensate", and now it has a second reason that owes nothing to a jury's judgement of quality: the instruction's referent is unfixed.
• What a practitioner can do instead, today: state the rule before the count — I am answering rhyme and root-play and not bare repetition, or the reverse — and expect the source's own readers to be using a different one. That is a smaller instruction than "compensate", and unlike "compensate" it can be checked.

The same run measures what the diagnostic is counting, on the side the gate does not touch. The 21 loci were also put to the seats source-blind, in the frozen ornamented English and in the translator's own plain alternative at the same span, paired by site:

• Where the Arabic is plain, the supplied device roughly doubles the reader's belief that the Arabic was ornamenting — 0.7500 with the device against 0.3750 without it, +0.3889 paired. Blind and with the device removed, readers land at 0.3750, close to the 0.2941 that readers of the Arabic itself endorse at those loci; with the device in, at 0.7500.
• Where the Arabic has a figure, removing the device changes the belief by −0.017 — because the plain English there still repeats what the Arabic repeats. Compensation's own sites are the ones where the English does least of the work.

So §7.14's warning is not merely about a reader failing to be informed: the supplied device manufactures the belief, and it manufactures it precisely where the source licenses nothing. That is the sharpest form the §7.7 diagnostic has taken.

Untested here (D-20260724-04): whether either rule matches what a reader of the Arabic responds to, as opposed to what they will name when asked; and whether a reader who can see the source behaves any differently, which this run's gate prevented it from asking. One chapter, 21 loci, three fixed models, one language pair, Tier D NOT PASSED; the seats are models, not readers, and three models agreeing is three models agreeing. The disagreement is evidence against the inventory rule, not proof that the seats' rule is right — classical rhetorical treatment also excludes rhyme by grammatical ending, and this page does not adjudicate. Limits in RS-20260816c §10.

7.16 The diagnostic counts marks; a translator defends them by naming a job — and on attitude the job he names is not the one the device does — 2026-08-16 (S201), RS-20260816g-device-function

⚠ THIS SECTION'S INSTRUCTION IS WITHDRAWN — see §7.20 (2026-08-20, S206). Step 2 put the same twelve stretches to an independent hand, and "write the passage without the device" turned out not to name one act: the second hand wrote a different device in. Item 1 below is withdrawn as an instruction; item 3 is withheld (this section's evidence for it cannot be re-checked, two of its six loci having been gated out); item 4's exception is restated as locus-specific; item 2 survives. Read §7.16 for the step-1 record and §7.20 for what the framework now says.

ARM-device-function step 1. Everything §7 tells a practitioner to do is a count: how many sound devices, how many marks of emphasis, how many pieces of unlicensed typography, against how many the source licensed. A practitioner asked to justify a device does not answer with a count. He answers with a job — it carries the manner, it carries the tone, it answers a figure. §7 has never said whether that answer is worth anything, and two results in the previous week said it might not be: RS-20260816f found a device redundant for the job its translator named and load-bearing for another, and §7.14 withdrew a structural account of a reader's inference that turned out to be backwards.

The test. 蒲松齡〈種梨〉 was translated whole into English by the lead under R04, and the frozen translator's log declared, before any design existed, a job for each of fourteen marked stretches from a closed vocabulary — MANNER, STANCE, ORNAMENT, NONE — with two further stretches marked and left untouched as a floor. A second version was then written, identical everywhere except inside the twelve declared markers, each replaced by the plainest rendering of the same events. Three seats, three replicates, one version per body, blind, three fixed yes/no probes at every marker.

1. The subtractive test works, and it is worth a practitioner's time. At nine loci out of nine it produced an answer clear of the instrument's floor (job-locus movement 0.864 against a sham floor of 0.278, measured on two stretches identical in both versions).

2. A translator can tell when a choice is doing nothing. The three loci he had declared inert — the three places where, revising, he changed the wording and could not say what the change bought — moved 0.0741 against 0.8642, exact P = 0.00455 over all 220 labellings. That is a real competence and it is worth saying so.

3. On manner and on ornament he was right, three for three and three for three. In great mouthfuls, craning his neck and staring, at an unhurried walk: the manner probe falls from 1.00 to 0.00 at all three and nothing else moves. Ten thousand eyes, in a flash, flowers; in a flash, fruit, chock, chock: the ornament probe falls from 1.00 to 0.00, 0.11 and 0.00.

4. On attitude he was wrong, three times in three, and the failure has a shape. At the three loci he declared STANCE, the attitude survives the subtraction at ceiling — 0.78 → 0.67, 0.78 → 1.00, 1.00 → 1.00. Readers agree the stretches carry attitude. They carry it in the plain version too. The countryman was stupid, and that is not surprising is contemptuous because of what it says; there is no plainer English for it that does not say it.

5. What those devices were actually contributing is the reader's belief about the source. At the same three loci the ornament probe falls by +0.333, +0.778 and +0.667. The seats name it unprompted: on the translator's version, "set address phrase", "idiom for begrudging facial expression", "rhetorical classical commentary phrasing"; on the plain version, the same seats at the same loci report the attitude and nothing else — "names their unwilling attitude", "narrator judges him and shrugs". And they are right about the Chinese: 於居士亦無大損, 怫然 and 蠢爾鄉人，又何足怪 are set forms. The reader recovered a property of the source that the translator, who chose the English for it, had filed under something else.

What v0.2 now says.

• The subtractive test is added to §7 as a diagnostic, in the same form as the count. Before you defend a device by naming what it conveys, write the passage without it and ask whether the passage conveys it anyway. Unlike "compensate", it can be run by one person on their own draft, and unlike the count it does not need a determinate denominator — the objection §7.15 and §10.7 raised against the two instructions before it.
• It has one stated exception and the exception is the useful part: it does not work on attitude. Where the device and the proposition are the same thing there is nothing to subtract, and the subtraction you can actually write removes something else. This is note (bpu) arriving from the other side — (bpu) found that a depictive device has no propositional constant; §7.16 finds that a stance device has no separable device. Two of the four channels §7 talks about are not manipulable in the way §7's advice presumes.
• And a warning that joins §7.14's. §7.14: a supplied sound device tells your reader the original was ornamenting, whether or not it was. §7.16 adds the case where the translator does not know he supplied one: he thinks he is carrying a tone and he is telling the reader the source had a set phrase. Here the source did. Nothing in his own account of the choice would have told him either way.

What is untested here (D-20260724-04): everything in item 5 about why — the seats' reasons are free text collected after the answers. And the central dependency is not answered: one hand wrote the translation, the loci, the labels and the plain alternatives, so the primary's hold licenses nothing on its own and is reported as consistent-with. The informative results are the three registered per-locus predictions that failed, which that dependency works against. One tale, 686 English words, one translator, three model seats, n = 9 job loci, Tier D NOT PASSED; the seats are models, not readers. Limits in RS-20260816g §7.

7.20 §7.16's subtractive test is WITHDRAWN as an instruction: asked to write the passage without the device, a second hand wrote a different device in — 2026-08-20 (S206), RS-20260820b-device-function-2

ARM-device-function step 2, and the arm closes resolved at 2 of 2. §7.16 (four days earlier) added the first §7 instruction that is not a count: before you defend a device by naming what it conveys, write the passage without it and ask whether the passage conveys it anyway. Its evidence had one hand writing the translation, the loci, the labels and the version without the device. Step 2 bought the one repair that mattered — an independent hand, blind to the labels and to any hypothesis, wrote the replacements — and held the translation, the loci, the labels, the three seats, the three probes, the caps, the temperature and the estimator byte-identical.

The instruction does not survive the repair, and it fails in a way that is about craft rather than about arithmetic. Given the translator's own rule in his own words — replace the marked stretch with the plainest English rendering of the same events; change nothing outside the stretch; add and remove no event — the second hand did not take the device out. It put another one in.

1. So the subtraction subtracted nothing where the translator said the device was working. Δmanner is 0.000 at all three MANNER loci — all eighteen bodies report manner in both versions, because the manner is in the adverb. The probe is not dead: it moves +1.000 elsewhere in the same bodies, and the separation criterion does not fire.

2. At the rhetorical question, eighteen bodies and two versions differ on nothing at all. Both are a rhetorical question calling the man a fool. A mechanical screen for a replacement identical to the original catches identical strings; it cannot catch a different string doing the same work.

3. What a practitioner has to take from this. Write the passage without the device is not a well-defined instruction. Two competent hands, same rule, same twelve stretches, produced incompatible subtractions, and the answer the test returns is decided by whichever of them wrote it. §7.16's instruction is withdrawn. What replaces it is narrower and checkable:

If you are going to test a device by writing the passage without it, write the replacement down and show it to someone else before you draw any conclusion from it — because your idea of "the same thing, said plainly" is doing more of the work than the device is. Where the two versions of "plain" disagree, that disagreement is the finding, and it is about your own reading of what the passage is doing, not about the device.

4. And the plain version is not the neutral one. At two loci the seats report more attitude in the replacement than in the original — Δstance −0.444 and −0.222 — and they say why, unprompted and in their own words: "greedily tells how and judges the eating", "greedily shows how and judges", against "describes only the way he ate" for the original. A depiction shows; the adverb that replaces it evaluates. Going plainer is not going unmarked; it swaps one mark for another, and on this material the plainer word is the more judging one. This is note (bpu) from a third side, and it is the sharpest practical form the point has taken: it is about a move translators make constantly and think of as neutral.

5. §7.16's exception on attitude is RESTATED as locus-specific, by a rule fixed before the numbers were seen. Of the three stance loci, one gave a clean rewrite, one gave none, and one was not rewritten at all. - 良朋乞米，則怫然 — they go sour in the face → they become angry: the attitude does not fall (0.667 → 0.778); the manner falls completely (+1.000) and the belief that the Chinese had a set form falls almost completely (+0.889). The seats name it: "idiom for scowling; stingy attitude" on the original, "names their angry attitude" on the replacement. 怫然 is a set form. §7.16's item 5 holds here, on an independently written replacement. - 於居士亦無大損 — no great loss to your worship → to you: the attitude falls (0.889 → 0.667), and one seat reports none at all in the replacement. Both gate seats independently flagged this pair — and only this pair of twelve — as changing "modality and social relation". - 蠢爾鄉人，又何足怪 — nothing was subtracted, so nothing is learned.

6. What survives of §7.16 unchanged. Item 2: a translator can tell when a choice is doing nothing. The three loci this translator declared inert, because revising he had changed the wording and could not say what the change bought, moved 0.1111 against 0.4603 at the job loci in this run's own data.

What is untested here (D-20260724-04): the reasons in item 4 are free text collected after the answers. What is withheld: §7.16's manner-and-ornament claim, which this run cannot speak to, because two of its six loci were gated out before dispatch under a rule fixed before any body was bought. What carries no evidential weight, by the design's own terms: the relabeling statistic — the labels were chosen from the visible wording for properties aligned with the probes, so no relabeling count is a test, whoever writes the replacements. Step 2 was constituted to identify step 1's primary and established instead that it cannot be identified from a label set the translator wrote; that is a loss and is reported as one. One tale, 686 English words, one translator, one writing hand, three model seats, n = 3, Tier D NOT PASSED. No cross-run comparison is made anywhere. Limits in RS-20260820b §7.

7.4 Q-e's obstacle is NAMED for the first time on the source-ward side — 2026-08-13 (S178)

RS-20260813g-register-quadrants, ARM-ennoblement closing resolved at 2 of 2. Every earlier entry under Q-e records a refusal and re-reasons it. This one records an obstacle, which is a different thing: a statement of what stands between the framework and the recommendation, measured rather than inferred.

Andersen's «Flipperne» (1848, Danish) in six English arms — three published hands and three lead renderings under no rule, under an ennoblement rule set (R25) and under a resistancy rule set (R08) — censused on sixteen Danish compounds frozen from the source before any English was tabulated, coded by two panel seats blind to arm under different label maps, agreeing at 0.9653 over 144 cells. A compound counts as carried only where both seats coded it carried.

The obstacle, in one sentence: a source-ward register recommendation has no published-practice precedent to borrow, because on this pair published practice does not occupy that quadrant at all. Three hands independently at three of sixteen, on different sites; one of them at zero of six fronting inversions; all three below a plain lead pass written under no rule whatsoever. A framework line of the form where the source compounds, compound would be a recommendation against what published translators do, not a distillation of it — and §5's traceability rule requires that to be said out loud before such a line could be written.

Two things this does NOT license. It says nothing about elevation: the run's Axis E voided on a dead body and where published hands sit above plain English is still unmeasured, so Q-e's original question — where the source marks low register, do X — is untouched. And it says nothing about readers: a formal difference of 16-to-3 is a difference available to be perceived, not one shown to be perceived. RS-20260813c §10 items 1–4 still describe the instrument that would answer both, and they are in wiki/backlog.md.

Evidenced on: Danish→English, one tale, one author. internal-judgment-only, provisional; Tier D remains NOT PASSED.

7.1 Q-e ANSWERED IN PART, and the refusal stands on a new reason — 2026-08-09 (S145)

RS-20260809g-device-cross, ARM-register-devices step 1. The two devices §7 could not separate were separated: R23 splits R22's bundled clause into a located-idiom switch and a nonstandard-spelling switch, and two generating hands wrote all four crossed cells as minimal revisions of their own placeless rendering, at the fourteen independently voted narration sites of RS-20260808e's Italian and Japanese cells.

Two things came back, and only one of them is licensed.

1. The register comparison is WITHHELD. P1a — what each device buys across a book's marked narration, zeros included — is −0.2083 in Italian and −0.5000 in Japanese, i.e. pointing at the spelling, consistently in both cells, in all four hand×cell blocks and under every leave-one-seat-out. The registered power gate fails in both cells and the figure is not read. A labelled post-hoc sensitivity that restores the one site a missing rating cost gives Japanese 6 non-tied sites at exact P = 0.09375 — still short of α = 0.05. Neither reading licenses it, and the bar was not moved.

2. P3 FAILS, and this is the licensed result. Nonstandard spelling is NOT a placeless device. Registered at loc(B) ≤ 0.25, the bar RS-20260808f's placeless arm met at 0.0588. Observed: 0.5714 and 0.5357 on two independent judges — and, in all four judge × cell blocks, the arm given only respelling is placed by the reader more often than the arm given every located idiom. Asked to name the marker, both judges name the respelling itself, and one attributes it: "an'" (US Southern/rural), on four separate arms.

The sentence §7 above rests on is false as stated on this evidence. A translator cannot "adopt phonetic spelling without relocating a book." An' is the commonest respelling English has, and it does not read as placeless lowness; it reads as somewhere. Refusing to put a Sicilian fishing village in Lancashire by writing it in eye-dialect puts it in Alabama instead.

So Q-e's refusal stands and its reason is replaced. It is no longer "this design cannot say which device costs the register." It is:

There is no cheap way down. The device that looked cheap — available to any translator, costing no relocation — is not placeless, and the register question that would rank the two devices is unresolved at a scale this project can reach. framework carries no register-carriage recommendation, and the obstacle has moved from the design to the language.

What a successor would need, named so it is not re-derived: the register comparison needs more sites per language cell than fourteen across two — the exact sign-flip test bottoms out at P = 0.0625 with five non-tied sites, and half of the device opportunities in this run produced no change at all. A cell of ~20 marked narration sites in one pair, with sites chosen for device availability rather than only for source markedness, is what would make P1a readable.

One thing Q-e must carry rather than rediscover, third measurement: reach is a property of the site's syntactic shape, not of the device or the language. The Italian sites are noun phrases — un bighellone di vent'anni, un codino marcio — with almost nothing a respelling can attach to, and the respelling reached 3 of 8; the Japanese sites are whole clauses and it reached 6 of 8. The respelling device attaches to function words and to the -ing ending, whose distribution has nothing to do with where a source marked its register.

7.2 Q-e's refusal stands a THIRD time, and the obstacle has moved off the design entirely — 2026-08-10 (S150)

RS-20260810c-register-reach, ARM-register-reach step 1. §7.1 said what a successor needed: a cell of ~20 marked narration sites in one pair, because the register comparison had died of power at 5 non-tied sites against a bar of 6. The cell was built — 30 sentences of Sōseki's 「坊っちゃん」 censused by three independent annotators, 8 arms, 3 seats, 720 of 720 rating cells returned, and G5 passed at 20 non-tied sites against a bar of 12. Power was not the problem.

The comparison is withheld a third time, and this time by the manipulation itself. Two independent hands, given explicit written permission to use regional slang, local idiom and class-marked grammar, changed their own English at 4 of 60 hand-sites — 0.067. The same hands, given permission to respell, changed it at 32 of 60 — 0.533. The located-idiom arm is byte-identical to the placeless baseline at 93% of hand-sites, so the check asking whether it is more located than the baseline returned −0.0167 and 0.0000 against a bar of +0.20.

So §7.1's reason is replaced, and the framework's refusal is now this. It is no longer "the comparison is underpowered"; it is:

The two ways down are not two options a translator weighs. On Japanese narration of this kind, under a minimal-revision procedure, one is taken half the time and the other is not taken at all. A recommendation ranking them would be ranking a device against a non-event, and the framework does not write one.

Three things the run licenses, all gate-free:

1. P3's failure reproduces on new material, new hands, new sites. loc(B) = 0.5833 and 0.5667 on the two judges, against ∅ at 0.1333 and 0.0667 — S145 measured 0.5714 and 0.5357. There is no cheap way down now rests on two independent runs, and the respelling that S145's judges placed in the rural American South is placed again here.
2. P2 holds at +0.6056. The unconstrained low pole moves the panel's register reading; the batch is not inert.
3. The site criterion admits 97 of 111 sentences — 0.874. The criterion was built to find islands of low register and this narration has no islands. Every figure in RS-20260810c is therefore about criterion-positive narration sentences of that span, not about marked sites at large, and §7's question changes shape for exactly the reason Q-b's did: the interesting quantity is not how much of the low register survives, but that in some books there is nothing else.

What a successor would need, named so it is not re-derived. The run cannot separate the located idiom is unavailable in this material from the located idiom is unreachable by minimal revision — an unexercised permission and an absent device look identical. The instrument that would separate them writes the +I arm from the source rather than from the placeless baseline, and pays the tie-rate cost the minimal-pair procedure was adopted to avoid. Two of the four located changes that did occur landed on the same two phrases the lead's frozen translator's log had named and refused (get up to, and the 尻を持ち込まれた site), which is why the distinction is worth paying for.

7.3 Q-e's refusal stands a FOURTH time, and the obstacle is now in the baseline — 2026-08-10 (S155)

RS-20260810z-idiom-reach, ARM-idiom-reach step 1. §7.2 said what a successor needed: an instrument that separates the located idiom is unavailable in this material from it is unreachable by minimal revision, by writing the +I arm from the source rather than from the placeless baseline. It was built — a full 2 × 2 of procedure × permission, 30 frozen sites, 14 arms per site, 840 of 840 rating cells, the same two judges as §7.1 and §7.2.

The primary failed and its null is WITHHELD: loc(Asrc) − loc(Nsrc) = +0.0667 / +0.1666 against +0.20 on both judges, and the instrument control the pre-run critic forced into the design — can these judges see a located idiom that a frozen translator's log documents? — failed on one of the two judges at +0.0667 against +0.20. v0.2 therefore does NOT say the located idiom is unavailable in this material, and §7.2's question is still open.

Three things the run licenses, and the first one changes what every earlier figure in §7 means.

1. THE PLACELESS BASELINE IS NOT PLACELESS. Two independent judges place the lead's R22 rendering — written under a rule whose entire content is no lexical item, idiom, or grammatical form that a competent reader would locate — at 0.0333 and 0.2000, and name, quoting the text: apologise, trodden, for a song, the state of me, had it brought home to me, give myself airs. Three are found in the frozen located lexicon mechanically, with no jury. At one site the placeless lead rendering says for a song where both models' placeless arms say for next to nothing — and the +I log lists for a song as a located choice taken under the permission.

Every quantity §7, §7.1 and §7.2 report is a difference measured from a floor that is itself located. And one of the markers is a spelling: apologise against apologize. There is no placeless orthography either. A translator writing English must choose a national spelling convention before choosing a single word, and five sessions of this programme have treated the −S switch as though standard spelling were one thing.

2. P1b reproduces exactly. loc(A) − loc(∅) = 0.0000 on both judges, against S150's −0.0167 / 0.0000, in a fresh batch on frozen text. §7.2's central number is not a batch artefact.

3. P3 reproduces a third time. loc(B) = 0.5500 / 0.5833, against 0.5714 / 0.5357 (S145) and 0.5833 / 0.5667 (S150). There is no cheap way down now rests on three independent runs.

And a hand effect that §7.2's C3 was already circling. Pooled across the two hands the permission buys +0.0667 / +0.1666; split, it is entirely one hand's. x-ai/grok-4.5 goes from 0.0000 / 0.0333 placeless to 0.1333 / 0.3000 permitted and 0.2333 / 0.5333 when told to prefer; qwen/qwen3.7-max at every dose is placed by L1 at zero of sixty. Swapping the hand at S150 did not fix what C3 was written for.

So Q-e's refusal stands, and its reason is replaced a fourth time. It is no longer "the two ways down are not two options a translator weighs"; it is:

The comparison has been asked from the wrong end. framework has spent four runs asking how much a located device adds to a placeless English, and the placeless English does not exist: two independent judges place it, on items a translator writing under a rule against them did not notice using, including a spelling that has no neutral form. No register-carriage recommendation is written, and the successor is not another dose of the device — it is a measurement of the floor.

What a successor would need, named so it is not re-derived. (a) A loc instrument that is shown to read standard-spelled located English, because half of this panel does not: G3b is the first check anyone has run and it failed on one of two judges. (b) A floor measurement — what share of ordinary competent English narration any two judges will place, with no register manipulation at all — which is the denominator every figure in §7 has been missing. (c) Hands selected for applying the manipulation, on a pre-dispatch check rather than after the fact.

7.4 §7.3's headline is WITHDRAWN, and the floor is measured — 2026-08-11 (S156)

RS-20260811-floor, ARM-idiom-reach step 2 as respecified. §7.3 asked for the floor and got it: eighteen texts, twelve of them published English narration by hands of documented nationality, measured on an orthographic nation-mark inventory frozen before any text was read; plus three judges on 46 passages and 20 one-letter minimal pairs. 18 of 18 bodies, 73 verifier checks, 0 failures.

§7.3's sentence "THE PLACELESS BASELINE IS NOT PLACELESS" is withdrawn as unsupported. T-botchan-R22-v1 carries one orthographic nation-mark in 1,969 words — 0.51 per 1,000. The least-marked of twelve published English texts carries 1.24; the same translator's English written under no rule at all carries 3.92. The rule bought a factor of 2.4 against published prose and 7.7 against its own author, and the single leak is apologise, a member of the one inventory family an independent pre-run critic contested as historically unsafe. Remove that family and R22 carries none.

What replaces the withdrawn sentence, stated at the strength the measurement supports.

1. A page can be nationally unplaced; a book cannot. The prediction that ordinary published English is located immediately — first mark inside 200 words — failed: the median first mark falls at word 626.5. What holds is that all twelve texts carry a mark, the last by word 1,363, and that where marks occur they are nearly perfectly one-sided — median purity 1.00, ten of twelve texts entirely British- or entirely American-spelled.
2. The denominator §7 was missing, in numbers. Published English narration: 1.24–5.33 orthographic nation-marks per 1,000 words; first mark at a median of word 627; purity 1.00. R22 is below the bottom of that range and R24 (1.45) is inside it.
3. §7.3's finding is not overturned — its scope is fixed. RS-20260810z's loc asked about country, region, class and period together. This run measures country only, and only orthographically, so trodden, for a song and the state of me are untouched by it. What is withdrawn is only the leap from apologise to there is no placeless English.
4. The reader is not the page. Asked which written convention a ~90-word passage follows, three seats answered CANNOT TELL on 0.480 of cells (CI 0.380–0.582), and when they committed, most of what they quoted was not language: Akaki Akakievitch, roubles, 'Stute Fish, sheepskin. Token concordance with the detector: 0.283.
5. But the spelling does the work when a reader meets it. One letter changed — colour→color, theatre→theater, grey→gray — flipped the verdict on 0.813 / 0.938 / 0.938 of sixteen minimal pairs, against an unchanged-duplicate baseline of 0.000 / 0.750 / 0.000. Two of three seats are clean; the third flips on unchanged text and is not usable.

And one thing about craft, which came out of the translating rather than the measuring. T-mikan-R06-v1 — Akutagawa 「蜜柑」 whole, single pass, no register rule — has a log frozen before any of this existed, recording thirteen nation-choices its translator noticed making. The detector found six, and every one of the six is in the log. The seven it missed are the ones no word list sees: a second-class carriage, the up train, sliding **backwards, the conductor (not guard), a **porter (not redcap), I should certainly have scolded, mandarins (not tangerines). Twelve British, one American, in one unrevised pass.

The translator's difficulty is not that the choice is invisible. It is that there is no way to decline it. §7's own regime R22 shows the choice can be very nearly refused at the level of spelling — but only by a rule that watches spelling, which no register rule in this project did until now, and not by a translator's ordinary attention.

Q-e's refusal still stands, and this run does not touch its reason. Nation-marking and register-marking are different channels. §7.2's question — is the located idiom unavailable in this material, or unreachable by the procedure? — is still open, is no longer carried by a live arm, and moves to §8 as an open question.

What a successor needs. (a) A period-calibrated inventory: the detector is contemporary and the corpus runs 1843–1930, and Poe's 1843 Philadelphia text is British-spelled in its first half and American in its second. (b) The edition question separated from the translator question — Hapgood's two New York publications read British on both instruments, and this design cannot say whether the hand or the house chose it. (c) Controls this run refused with a reason: a sham edit and a mark-masked passage (E-20260811-floor/critic.md, A9).

7.5 The reader-side half of §7.4, measured — 2026-08-11 (S157), RS-20260811b-realia-channel

§7.4 measured the page. This measures what a judge does with it, and the two channels a translated page is located by turn out to behave differently.

E-20260811b crossed realia kept / muted with British / American orthography, within passage, on seven translated passages, so that author, subject, period, difficulty and recognition are constants and only the manipulated spans move.

1. Deleting the entire source-culture world from a page does not change these seats' verdict about whose English it is. +0.048 (CI −0.069 to +0.164, exact P = 0.3125, n = 7), against a registered ≥ +0.15 and P < 0.05. This is a failure to detect and is NOT reported as separability — the pre-run critic forced the equivalence margin (±0.10) before the run, and the interval does not fit inside it.
2. Changing the spelling does not change where the story is. Setting answers moved on 2 of 42 orthography pairs (0.048, inside the registered margin), and both moves were an abstention resolving rather than one country becoming another.
3. §7.4's own worry is bounded rather than dismissed. That run's judges quoted roubles and Akaki Akakievitch when asked about the English, and its G1b failed at 0.283. Under this controlled test the same behaviour appears — 13.5% of prose-question cues are realia — and does not change the verdict. The instruction fails at the level of what a seat reports reading, not at the level of what it concludes.
4. The finding a translator should take from this is the one nobody designed for. All twelve realia cues are three words — halfpennies, smock, councillor — and every one is the translator's own Anglicisation of a foreign thing. Garnett's Russian peasants gamble with halfpennies; three seats pointed at that word and said British, on both the British-spelled and the American-spelled form. The retained foreign words — taiga, yamen, Sanzu-no-Kawa, the Dragon Throne — were cited 118 times for where is this set and not once for whose English is this.

So the channel that leaks is domestication, not foreignisation. A translator who keeps the source's word is not, on this evidence, buying a stronger impression that the prose is nationally placed; a translator who reaches for the nearest English institution may be.

What §7 may now say, and what it may not. It may say that on these seven passages and these three seats, the two locating channels did not measurably interfere at the level of verdicts. It may not say the channels are independent, that the effect is zero, or anything at all about human readers — amendments A14, A18 and A21 of E-20260811b, all forced by the pre-run critic and all recorded in critic.md.

7.6 The three roads out of a culture-bound word, and the trade none of them escapes — 2026-08-11 (S162), RS-20260811h-domestication-channel

§7.5 ended on a finding nobody had designed for: the realia that leaked into the judgement of the prose were the translator's own Anglicisations — halfpennies, smock, councillor — and never a retained foreign word. This tested it, and §7 can now say what the roads trade.

E-20260811h built eight translated passages into four forms differing only inside 49 declared culture-bound sites — TRA (the source word carried over) · DOM (the nearest article of the English domestic world) · NEU (a location-free English phrase) · an edit-matched SHAM — and put them to three seats as forced pairwise comparisons. Six passages are a fresh lead rendering of Lazarević's «Први пут с оцем на јутрење» (1879); two are E-20260811b's Garnett and Field, so the word that generated the conjecture is in the corpus in the hand that wrote it. Every form is American-spelled by one mechanical map, so nothing here is a spelling.

1. Domesticating beats transferring on "whose English is this" at 24 of 24, 8 of 8 passages (exact one-sided sign P = 0.0039), and beats the location-free rendering at 48 of 48. This contrast, C6, exists because the pre-run critic pointed out that the asymmetry the whole claim rests on was never being tested directly.
2. Transferring places the prose the other way. TRA over NEU on the same question: 6 of 40 = 0.150, a registered equivalence interval failed low. A page with dukat and Đurđevdan in it reads less British than the same page with gold piece and the spring hiring-day.
3. It is nationality, not archaism. 89.6% and 95.8% of the deciding cues land on sites a translator's log had marked specifically-British before the design existed; the edit-matched sham sits at 0.381; and the same seats call the transferred words foreign at 24 of 24 when asked about the world.
4. One nation-marked word in 347 words flipped 6 of 6 judgments. The effect is at ceiling on every passage, which means its size is not measured — only its direction.
5. And the confound check did not clear. Asked which version is more clearly set outside the English-speaking world, the seats chose the domesticated page at 1 of 17. Domestication moves the world as well as the prose.

What §7 may now say. The three roads trade world-carriage against prose-placement, and none of them gives both: transfer places the world and pushes the prose away from the target's own nation; domestication pulls both toward it; neutralisation is the only road that leaves the prose unplaced, and it buys that by deleting the world. A translator reaching for the nearest English institution in order to read naturally is not producing neutral English — they are producing English of a particular country, and a less foreign story with it.

What it may not say. Nothing about magnitude (everything is at ceiling), nothing about human readers (three seats, an instructed task), nothing about moderate domestication (DOM is the ceiling: every item at once), and nothing that overturns §7.5 — that absolute-verdict null still stands, and the two together say the difference is perfectly detectable side by side and did not move a verdict given alone.

7.7 Punctuation nobody chose, and the one instruction §7 will carry — 2026-08-12 (S163), RS-20260812-unlicensed-typography

§7.6 measured what a translator's substitutions do to a reader. This measures the smallest thing a translator changes — the marks — and asks who is changing them. The demand side is RS-20260811c: three blind seats reading one English rendering and answering about the Swedish scored 0.0000 on the two typographic lures, every seat and every segment, at a mean confidence of 5.887 of 6 on their wrong answers, and the two arms that isolate the effect were certified propositionally identical 6 of 6 by an independent parity seat. A punctuation mark in a translation is read as the author's, and read hardest when it is not.

E-20260812 asked who supplies it. Two censuses, arithmetic only, no jury, one API call (the pre-run critic, which returned NEEDS-REDESIGN with seven BLOCKING findings, all seven accepted — the primary measure in this row is the one the critic specified, not the one the design froze).

1. A careful single pass supplies almost nothing deliberately. Garshin's «Очень коротенький роман» against the lead's R06 rendering, 32 paragraphs aligned to 32: 58 of 58 exclamation marks, question marks and suspension points land in the aligned paragraph and not one of the three kinds is added. 94.25% of every mark the English reader sees is licensed by a source mark of the same kind in the same place.
2. What does move, moves by convention. All fifteen omissions in that translation are dashes and all fifteen sit in the fourteen paragraphs where the source has direct speech: they are the Russian dialogue dash becoming an English quotation mark. Zero omissions in the eighteen narrative paragraphs.
3. And the translator predicted the direction wrong. A prediction frozen in git before any count existed said the dashes would rise by about a third; they fell by 48%. He was exactly right — 0.00% error, three times over — about the marks he kept.
4. A whole class can vanish without anyone deciding. Gogol's «Шинель» carries 21 suspension points. Hapgood 1886 has none and Field 1916 has none; Kassner's German keeps 19. Whether that is two translators' choice or two publishers' house style cannot be recovered from the texts — and to the reader it makes no difference.

What §7 declines to recommend, explicitly. Not "preserve the source's punctuation". The demand evidence is three prompted LLM seats on an instructed task with Tier D NOT PASSED; the strongest known lure, italics, is unmeasurable across these transcriptions; retention here is measured at paragraph grain and is therefore an upper bound; and the one case of wholesale deletion may be a publisher's rule rather than a translator's. Nothing licenses a quality rule about marks.

What §7 does adopt, and it is v0.2's first instruction a translator can execute unaided: count the source's expressive marks, count your own, and look at the difference — because the difference is where the changes you did not decide are. The warrant is not a jury. It is the failed prediction in item 3: a translator who had just finished the text, asked to say what his own punctuation had done, got the sign wrong on the only mark that moved. That is arithmetic, it needs no calibration, and it is the cheapest diagnostic in this framework. It diagnoses; it does not prescribe. What to do about a difference the count surfaces is not evidenced here, and §7 says nothing about it.

7.8 The diagnostic extends to emphasis, and emphasis is where it bites hardest (2026-08-12, S164)

§7.7 wrote emphasis out of the diagnostic in the same breath as it adopted it — "the strongest known lure, italics, is unmeasurable across these transcriptions". It is now measured, on materials where three witnesses mark emphasis machine-readably: A-poe-emphasis-carriage / RS-20260812b-emphasis-carriage, seven tales of Poe against Baudelaire's French of 1857 and Sasaki's Japanese, 74 source emphasis spans and 244 target ones, span-level.

The count changes from a small correction to a large one. For punctuation, 94.25% of the marks the reader saw were licensed by the source. For emphasis it is about a third — 0.3333 in Baudelaire, 0.3305 in Sasaki, two hands, two centuries, two different emphasis devices, the same number. Excluding the one tale with prior exposure, 0.2453 and 0.2245.

Three things a translator can act on, and one they cannot.

1. Your stress marks are probably fine; it is the ones you did not think of as emphasis that flood. Both published hands carry the source's stress italics at 0.833 and 0.944. What arrives unlicensed arrives beside them: titles italicised by house rule, foreign or low-register words marked because they are foreign in your language, and stress the source simply does not have.
2. A foreign-word italic is carried by function, not by mark. Poe's twenty-five foreign-word italics: Baudelaire keeps the eight still foreign in French and drops every French one; Sasaki keeps none of the twenty-five, because katakana already says foreign. Dropping a mark whose job your writing system already does is not a loss, and your count will call it one — which is why §7.7's instruction diagnoses rather than prescribes.
3. Inventory the source's emphasis devices before counting, not just its italics. Poe also uses small capitals; both hands convert some into their own emphasis device, because neither language has small capitals for that job. A count restricted to italics scores those as additions.

What v0.2 still declines. No rule about preserving emphasis. The demand evidence remains three prompted LLM seats with Tier D NOT PASSED; the supply evidence is a correspondence among three modern transcriptions and cannot separate a translator's mark from a compositor's; and the correspondence coding is single-coded. The instruction stays what §7.7 made it — count, and look at the difference — with emphasis now inside it.

7.8a The two channels are alternatives: where the foreign word survives, the mark does not (2026-08-12, S169)

§7.8 item 2 said a foreign-word italic is carried by function rather than by mark, and read the mechanism off two hands. A third hand sharpens it into something a translator can use. RS-20260812g-foreign-italic put Konstantin Balmont's Russian Poe (Скорпион, 1901–1912) against the identical 74 source spans, span-level, with the correspondence calls on the class that matters independently blind-coded.

On the nineteen measurable foreign-marked spans, Balmont keeps the mark on 7 of the 10 he translates into Cyrillic and on 0 of the 9 he keeps in Latin letters — while carrying 35 of 35 of Poe's ordinary stress italics, so the drop is a decision about foreignness and not a distaste for italics.

The instruction that follows, and it is still a diagnostic, not a rule. Ask, per mark, which channel is carrying the foreignness in your sentence.

• If you keep the source's foreign word as a foreign word — Latin script inside Cyrillic, katakana inside Japanese — the italic is saying a second time what the alphabet already says. Dropping it loses nothing your reader can perceive, and your count will score it as a loss.
• If you naturalise the word, the italic is no longer decoration: it is the only surviving trace that the author was pointing at that word. Dropping it there is a real loss and the count will not tell you, because the count is the same either way.

What v0.2 declines here. The evidence is one hand and one language pair, exploratory rather than confirmatory (the corpus was counted before the predictions were written), and nothing separates "Russian does this" from "Balmont does this". No jury; Tier D NOT PASSED.

And one caution about the numbers in §7.8, from the same run. Balmont's licensed rate is 0.8723 in the 1901 volume and 0.1034 in the 1906 one — a factor of 8.4 splitting exactly on which printed volume the transcription came from, with stress retention at 1.000 in both, so it is not simply lost markup. A single "licensed" figure for one hand can be an average over volumes that have nothing to do with each other. Report per copy-text, or say that you did not.

7.9 The middle of the road, measured: the prose is a slope and the world is not measured as one — 2026-08-12 (S167), RS-20260812e-dose

§7.6 described a translator nobody is. Its domesticating arm replaces every culture-bound item in a passage at once, and §8 below has carried the consequence since S162: a translator who domesticates four sites in forty is not described. Every practising translator is that translator. This measures the road between the two ends.

The design. Ten translated windows — four new, cut from the lead's R06 rendering of the whole of Neruda's «Hastrman» (1878), Czech, frozen with its 33-site culture-bound table before this design existed; five from §7.6's Serbian windows; one a published English hand — each built into a nested dose ladder over a seeded permutation of its own declared sites: none domesticated, one, half, all, plus a control differing from the one-item form at exactly one site, the English domestic article against a location-free English phrase. 324 bodies, 0 dead, 0 truncated.

1. Nothing saturates. One item beats none at 28 of 28; half beats one at 30 of 30; all beats half at 30 of 30; all beats one at 60 of 60 — 10 of 10 windows on each, sign P = 0.00098. The registered equivalence window [0.35, 0.65] on the primary was reachable at n = 60 and was not reached. Every further Anglicised item is detected again against the page that has fewer.
2. And one item is already enough. The bottom rung is 1.000, the same as the full-substitution control. In one window the two pages differ by one word in 245 — "Three zlaté!" against "Three guineas!", with the line two above still reading three zlaté for it, in dvacetníky — and all three seats chose the second as the more British page, in every window.
3. It is the domestic article, not the missing foreign one. The same-site control — guineas against gold pieces, greengrocer against shopkeeper, beaver against tall hat — goes 56 of 58 = 0.966, 90% CI [0.895, 0.994], 10 of 10 windows. §7.6's rival mechanism, that removing a foreign word is what moves the verdict, does not explain the one-item effect.
4. On the world, one item is already total against the untouched page: the fully transferred version is chosen as the more clearly foreign-set at 30 of 30 against the one-item version, and at 30 of 30 against the fully domesticated one.

What §7 may now say. The three roads of §7.6 are travelled item by item, and what a translator does not Anglicise, they keep: partial domestication is a real intermediate position on the prose channel. It is not a discount on the other one. There is no measured dose at which the setting is left alone — a single English domestic article, in a page carrying four to sixteen source words, is already enough for these seats to name the untouched page as the foreign-set one. A translator taking the middle road should expect to keep some of the source-oriented texture of the English and should not expect the story's foreignness to be protected in proportion.

What it may not say. Nothing about magnitude — every prose contrast is at 1.000, and the only quantity in the run is a post-hoc confidence ladder (2.54 at one item → 3.00 at all of them) that licenses nothing. Nothing about human readers. And nothing about whether the world cost grows: the design compared both doses against the untouched page and never against each other, which is its own gap and is ARM-dose step 2's first candidate.

7.10 The other road out, measured: what you put in the gap still decides where the story is — 2026-08-13 (S172), RS-20260813a-world-dose

§7.9 above says there is no measured dose at which the setting is left alone, and that sentence invites a conclusion it does not support — that the story's foreignness is all-or-nothing, spent at the first Anglicised word, so the middle road buys nothing on this channel. This run was designed to test the sentence and had to be rebuilt before dispatch, because the contrast it registered was not a measurement: comparing a page carrying six source words against the same page carrying none, and asking which is more clearly set outside the English-speaking world, measures the source words. The independent pre-run critic said so and the design was changed (RS-20260813a §5, note (bmx)).

What was measured instead. A sixth form in which every culture-bound site is rendered in location-free English — so that the fully Anglicised page and the fully location-free page carry zero source words each and differ only in the kind of English that replaced them. On the Greek shop window: a pinch of snuff … his breeches … his nightcap … Mr. Margaritis's counting-house against a pinch of powdered tobacco … his baggy trousers … his woolen cap … Margaritis's shop. Fifteen windows, four source languages, 252 judgments.

1. The kind of English still decides it. The location-free page is chosen as the more clearly foreign-set one at 42 of 49 cue-attributed judgments across every site, and at 30 of 33 when only one site differs. Both registered intervals fall entirely below the equivalence boundary. So the third road is not a nothing: after the source word is gone, reaching for a location-free phrase rather than an English domestic article still holds the story further from England.
2. And the evidence is weaker than that sentence sounds. The design's own inferential unit is the window, and on the ten primary windows the sign test returns P = 0.090 and P = 0.145 — the direction is nearly unanimous, the window-level test does not reach the conventional bar, and nine usable windows cannot produce it. Reported as a direction with a pooled effect, not as a window-level result.
3. It reverses at proper names, and a translator should expect that. The one window that goes the other way does so on Prague's Malá Strana rendered "the Lesser Town" against the location-free "this quarter". At a name, Anglicising is a transparent calque — it Englishes the words and keeps the place — while the location-free move is a deletion that names nowhere. The two roads swap burdens at toponyms, and the guidance in 1 does not extend to them.
4. Where a place name survives untouched, the judgment stops moving at all. Refusals ran an order of magnitude above the prose channel on the same materials (0.087 / 0.212 / 0.263 by seat against 0.010 / 0.020 / 0.010), and of 45 refusals 14 quoted "Prague". Post-hoc and descriptive, but it points the same way as 3: the names do work the furniture cannot undo.
5. §7.6's core result replicates in a fourth language pair — the fully domesticated page reads as the more British-written one at 14 of 14, 5 of 5 windows, on a modern-Greek source translated the same day.

What §7 may now say to a translator. After §7.9's slope on the prose channel, this: the choice between an English domestic article and a location-free phrase is not only a choice about how English your prose sounds — it is also a choice about where your story appears to be set, and it stays live at a single word. And, immediately after it, the boundary: this does not hold for place names. Calquing a toponym keeps the place; neutralising it deletes the place.

What §7 may NOT say. Nothing about magnitude. Nothing about human readers — three instructed LLM seats, temperature 0, a jury that has not passed its calibration gate. Nothing that distinguishes reads as belonging elsewhere from reads as translated: the location-free renderings are measurably the longer ones (0.53 words per site on average, +2.00% over the corpus) and the phrases the seats quoted as their reasons are exactly those periphrases. And nothing about whether the world cost grows with the number of items domesticated — that question is now recorded as one this instrument cannot ask, rather than as an open gap.

7.13 Q-e's reader-side leg is refused again, and the craft leg S189 opened is WITHDRAWN — 2026-08-15 (S194), RS-20260815e-fluent-carriage-2

ARM-fluent-carriage closes resolved at 2 of 2. §7 has refused, three times, to recommend source-ward moves on the ground that readers register them as source-driven. This entry says why the refusal must stand and takes back the one thing its predecessor gave.

The instrument was rebuilt until it could be read. Step 1's arms failed an independent content-parity control on 0 of 7 segments and every carriage figure it produced was withheld. Step 2 dropped the three device classes that control had convicted, had a blind reader price the remaining site census (25 of 33 — the mimetic class at 1 of 6), rebuilt the oddity arm, and passed the same control on 7 of 7 with 5 of 5 planted errors caught. It is the first carriage manipulation in this project whose figures may be read at all.

What it found, in three sentences a handbook has to live with.

1. An arm that carries nothing and is merely badly written is chosen as the source-follower over a clean flattening in 20 of 21 cells. The four false-positive figures §7 already carries now have a forced-choice replication with content parity established.
2. Set carriage against non-fluency and readers do not agree — indeterminate at 11 of 21 here and 13 of 21 at step 1, the seats disagreeing with each other both times.
3. Give the reader the original and the picture inverts: the same pair, same seats, source printed above, goes to the carrying arm at 20 of 21, both reversing seats moving the same way.

So Q-e's reader-side leg is not restored, and the reason is now specific rather than cautionary: a source-blind reader's judgment of source-following is substantially a judgment of how translated the English sounds. A recommendation may not rest on it. A source-visible reading is a different measurement and this run did not register a bar for it.

And the craft leg is withdrawn. RS-20260815b §6b licensed the sentence "carrying a source's form does not require paying in fluency" on a quality-parity gate passing at 8 of 21 (P = 0.383). On the cleaner manipulation the same gate runs the other way: readers prefer the flattened arm as English at 15 of 21, two-sided exact P = 0.0784. That rejects the run's registered parity interval and is the wrong direction for the claim. It is a lean, not a demonstration — it does not reach conventional significance — so what replaces the withdrawn sentence is not its converse but silence: whether carriage is free is unmeasured, and the one clean attempt leans against it.

No recommendation is added. R1's text is unchanged. What §8 Q-e gains is a third refusal with a measured reason, and what it loses is a craft claim that should not have been deposited on a gate that could only fail to detect a difference.

7.17 §7.15's denominator is determinate after all — and the thing that was rule-dependent was one clause of this handbook's own inventory rule, which is now struck — 2026-08-16 (S202), RS-20260816h-target-set

ARM-invented-ornament step 2, and the arm closes. §7.15 read RS-20260816c's failed gate as showing that "the places where the source has a sound figure" is not a determinate set, because two defensible rules pick out different ones. On fresh material with the question put in the translator's own terms, that is too strong, and this section narrows it.

The test. «باب اللبؤة والإسوار والشغبر» — a fourth chapter of «كليلة ودمنة», 599 Arabic words — was enumerated twice before any English was written: once under this project's standing inventory rule (Rule A: matched shape or a repeated word, with rhyme by an enclitic pronoun excluded), once under the rule RS-20260816c's seats were observed applying (Rule B: rhyme however produced, plus a root in a changed shape, and nothing else). The two rules agreed on 5 of 24 places. Then three seats were shown the Arabic alone, one locus marked, and asked the translator's question: would a translator be right to treat this place as a sound effect of the original that his English ought to answer?

The registered primary — that the readers would side with Rule B against Rule A by ≥ +0.40 — FAILED at +0.2778. They side with almost everything:

The level clause held at both ceilings — the places both rules admit are owed at 1.0000, the places neither admits at 0.0000, 24 of 24 bodies — so the jury discriminates; it simply does not decline figures.

Three things go into this handbook, and the first is a repair to its own instrument.

1. Strike exclusion (a) of the inventory rule. Rhyme carried by a shared grammatical ending or an identical enclitic pronoun was excluded by name from this project's figure inventory, on the classical ground that it is grammatical rather than chosen. Every place excluded by that clause and put to the seats is owed, 9 of 9 bodies, named as saj' at 3 of 3 seats at two of the three loci — and it replicates RS-20260816c's one disagreement that did replicate, in a different enclitic. It is the commonest sound effect in this prose and the rule was throwing it away. The four chapters of this work already enumerated under exclusion (a) have figure counts that are therefore low.
2. Answer the union of a form rule and a sound rule, not one of them. Rule B is a strict subset of what these readers owe — it never over-claims and it misses 8 of the 19 places they do owe. Rule A contains 13 of the 19 and over-claims at 4. The union contains all 19 and 4 spare. Answering the union misses nothing and over-answers at about one place in six, for about 4.5% in length — measured, on this chapter, by the rendering that did it. Amended 2026-08-16 (S203) by §7.18: say, when issuing this instruction, that the one published English hand translating from the Arabic answers 15% of these loci and 0% of the matched-shape ones. A practitioner following this instruction is doing something the available precedent does not do.
3. untested: a repetition close by is owed; a repetition far apart may not be. The one class that split, split by distance and not by kind — العدل … العدل across two words is owed at 1.000, أكل … أكل across the chapter's widest span at 0.000, with the seats naming the repetition in both. Unregistered, read off seven loci, and the successor's business.

What §7.15 keeps and what it gives up. It keeps its practical instruction — say which rule fixes your target set — because the two rules do differ at four fifths of the licensed places, and a translator following one answers where the other says plain. It gives up the stronger claim that the target set is indeterminate: on this chapter the readers' set is determinate (19 of 23 figure loci, 0 of 5 non-figure loci, unanimous at 24 of 28), and what was rule-dependent was a clause of §7.15's own instrument. §7.15 and §7.17 must be read together, in that order.

Limits. Three models, 28 loci, one chapter, one work, one language pair, Tier D NOT PASSED; the seats are models, not readers of Arabic, and three models agreeing is three models agreeing. One hand wrote both rules, both inventories and the class map; the seats' own labels corroborate it at 24 of 28 and that is a weak check, not a repair. The primary failed and no claim is made that these readers apply Rule B. Limits in RS-20260816h §10.

7.18 §7's Arabic instructions now have a published human hand behind them — it answers one figure in seven, no matched shape at all, and its 1818 preface says so in advance — 2026-08-16 (S203), RS-20260816j-published-figure

ARM-published-figure step 1, and the first Tier 1 anchor this project has with Arabic as source (A-knatchbull-kalila). §7.15, §7.16 and §7.17 instruct a practitioner about the sound figures of an Arabic source, and until now every one of those instructions rested on this project's own hands — the lead's renderings and three model seats. §7.17 went further and struck a clause of the standing inventory rule on three models owing a class 9 of 9. This section puts a published translator under all of it.

The pair. Wyndham Knatchbull, Kalila and Dimna, or The Fables of Bidpai. Translated from the Arabic (Oxford, 1819), read whole against «كليلة ودمنة» at the frozen figure loci of two chapters — 54 in «باب القرد والغيلم», rendered by the lead under R38 and frozen before the comparator was opened, and 24 in «باب اللبؤة والإسوار والشغبر», of which three readers of the Arabic endorsed 19 at RS-20260816h. Seventy-two codable loci; six lost to a scan defect.

Three things go into this handbook.

1. State, when recommending that a translator answer the source's figures, that published practice does not. §7.17 instruction 2 tells a practitioner to answer the union of a form rule and a sound rule — about one place in six more than the sound rule alone, for 4.4% in length. The one published English hand from the Arabic answers 15% of the loci and none of the matched shapes. That is not an argument against the instruction; it is a fact the instruction should be issued alongside, because a practitioner who follows it is doing something the available precedent does not do. §7.17 instruction 2 is amended to say so in its own text.
2. §7.17 instruction 1 is corroborated by a non-model hand, at 3 of 11 and not at 2 of 2. Knatchbull answers both enclitic chains in «اللبؤة» — six Arabic clause-ends on ‑humā become discovered them, and shot at them, and killed them, and tore off their skins — and one of nine in «القرد». The difference is a fact about English, not about him: the answered chains are runs of transitive verbs sharing one object, where English puts a pronoun at the clause-end whether anyone is answering anything or not. Cite the 3 of 11, never the 2 of 2.
3. A translator's declared programme predicted the census. Knatchbull's 1818 preface says a literal translation is the wrong side of a beautiful piece of tapestry, that too close an imitation would be offensive to a modern ear, and that it was impossible to express the sententious brevity of the Arabic with strict fidelity. Muwāzana is the sententious brevity, and it is the class at 0 of 18. The shelf's second declared programme after A-morri-botchan, and the first one that turned out to be right about its own cost.

A fourth thing this grounds outside §7. The preface is a fully-formed statement of the fluency norm, in English, in 1818, about a non-European source — a primary instance of what S-venuti-invisibility describes, from inside the period rather than about it. Usable wherever that source is cited.

Limits. One translator, one work, one recension against another (de Sacy 1816 against Būlāq 1937 — three of the seven absent loci are recension divergences, not cuts), two chapters, 1819 English. All 72 three-way calls are the lead's, unchecked, made by the party holding the hypothesis — the finding's weakest joint, and ARM-published-figure step 2's whole business. The coding rule was made deliberately permissive on the answered side so that its bias runs against this finding. Tier D NOT PASSED; no jury was involved and nothing here is a quality claim. internal-judgment-only. Limits in RS-20260816j §7.

7.19 A device the target language has no category for: the English phonaestheme at a Japanese mimetic is judged faithful with the mimetic in the sentence and an invention without it — 2026-08-17 (S204), RS-20260817e-mimetic-reading

⚠ THIS SECTION'S INSTRUCTION IS WITHDRAWN — see §7.21 (2026-08-20, S207). The subtractive test §7.19 recommended was run again against a fair subtraction — the mimetic replaced by the plainest grammatical Japanese phrasing of the same event, with the substitute written independently by two hands — and the enacting English flipped from adds nothing to ADDS at 1 of 7 primary sites, not 7 of 9. The mimetic-present baseline this subsection stands on also did not itself reproduce: 5 of 7 on the aggregate continuity check and 4 of 7 per item, for reasons that have nothing to do with the mimetic — the seats' flags moved because of material in the English frame (「in her hands」, "bath was ready") or a shifted attention to an omitted clause. What replaces §7.19's instruction is narrower and is stated in §7.21. Read §7.19 for the step-1 record and §7.21 for what the framework now says.

The first §7 subsection about a device English has no grammatical category for at all, and the section ARM-mimetic-carriage was constituted to write. Japanese 擬音語・擬態語 — reduplicated bimoraic kana, or a bimoraic form with 〜り/〜と — are a closed morphological class with no English counterpart. RS-20260816e established two things about them: the class is recoverable (two readers who never saw an English word listed the translator's sites at 14 of 14 and 13 of 14 on chapter 2, and 10 of 10 and 8 of 10 on chapter 3), and an enacting and a stating English rendering of the same site are not paraphrases — they assert different things, so the marking cannot be priced against a constant.

What was evidenced by S204's design, and is now qualified by S207. §7.16 tells a practitioner to write the passage without the device and ask whether it does the job anyway. This subsection ran the same subtraction on the source, with the English held byte-identical: three seats judged the translator's two candidate renderings against the Japanese sentence, and against the same sentence with the mimetic deleted.

Instruction (JA→EN, 擬音語・擬態語) — WITHDRAWN 2026-08-20 (S207), see §7.21. At a Japanese mimetic site, an English phonaestheme — a sound-symbolic verb, a phonaesthemic adjective, a reduplication — is licensed by the mimetic and not by the event. Reaching for one where the source has a mimetic is, on this evidence, an accuracy move; reaching for one where the source has none is an addition, and reads as one. The practitioner's test is the subtraction: cover the mimetic in the source and ask whether your English word still has anything to answer to.

Evidenced on: 夏目漱石「坊っちゃん」chapters 2 and 3 translated whole by the lead (T-botchan-ch2-R06-v1, T-botchan-ch3-R06-v1), 15 mimetic sites, 9 admitting a clean deletion. With the mimetic present the enacting rendering is judged to add nothing at 13 of 14 resolved sites; deleted, the same words become an addition at 7 of 9 pairs, with no pair moving the other way — exact two-sided sign test p = 0.016, and p = 0.125 under the stricter of the two drop-rules the frozen design contains (RS-20260817e §6, a defect of the design, reported both ways). The direction — 7–0 and 4–0 — is the same under either rule.

The second half of the instruction, and it is a warning. At these sites the stating rendering is the one more often judged to lose something: it adds at 2 sites and omits at 4, against the enacting rendering's 1 and 0. A practitioner who reaches for the plain adverb because the phonaestheme feels like ornament is, on this evidence, choosing the arm that loses more.

What is NOT evidenced, and it is a great deal.

1. Three named model seats (P2 P3 QR), on 24 curated items from one novel. Charter §4 forbids treating panel agreement as validation; nothing here is a claim about Japanese, only about what these readers said. Tier D NOT PASSED; no jury and no quality claim is involved.
2. A deleted sentence is not a sentence Sōseki wrote, and a seat may be answering the mutilation rather than the absence of the property. This is the live threat and no control here addresses it.
3. Nine pairs, six under the strict rule.
4. The sound/manner split is not tested. 6 clean sound sites against 2 clean manner ones is not a contrast, and the by-class counts are refused as evidence.
5. Where no phonaestheme can be built, this instruction says nothing. Nine of the 24 sites had no enacting rendering the translator could construct; a blind second hand asked for alternatives at all nine returned 54 candidates, every one of them a statement — reported in RS-20260817e §7 and registered in advance as carrying no evidentiary force, because a seat declining to produce a phonaestheme does not show English lacks one.

internal-judgment-only. Full limits in RS-20260817e §9.

7.21 §7.19's subtractive test is WITHDRAWN as an instruction: with the mutilation gone, the flip is gone too — and the baseline was reading the English frame — 2026-08-20 (S207), RS-20260820c-mimetic-subtraction

ARM-mimetic-subtraction step 1, and the arm closes resolved at 1 of 2. §7.19 was four days old and it acknowledged, in its own §9 item 3, that a deletion is not a sentence Sōseki wrote and that a reader might be answering the mutilation rather than the absence of the property. It also (§9 item 5) allowed that the chapter-2 items were built for a different design and carried item-level confounds. This section closes both.

What was run. The same nine sites RS-20260817e used, and the same three seats (P2 P3 QR) reading the same two English strings (the enacting mk and the stating pl renderings) byte-identical, against the Japanese sentence with the mimetic replaced by the plainest grammatical Japanese phrasing of the same event, carrying no phonaesthetic form. Two independent hands — the lead and P1, blind — wrote the substitutes; they agreed at 8 of 9 sites under the frozen G3 rubric before dispatch (note bqm). Two sites (M08, M12) were excluded from the primary before dispatch because a colloquial intensifier or a second untouched mimetic-derived verb stranded on the substituted sentence — recreating the mutilation the run existed to avoid. Primary denominator: 7.

The primary FAILS. On the seven primary sites, the enacting arm's majority ADDS flag moved: 1 forward (no→yes), 1 back (yes→no), 5 unchanged. The registered directional criterion (≥ 5 forward AND ≤ 1 back) is failed on both clauses of the AND. Exact two-sided sign test on the two discordant pairs, p = 1.000, reported descriptively.

The baseline did not reproduce either. The mimetic-present cell of this run matched S204 at only 5 of 7 on the aggregate check and 4 of 7 per item, so under the frozen F4 gate the between-run contrast (deletion at S204 vs substitution here) is void as a controlled comparison — only the within-run null is clean. And the three drift sites drift for reasons that have nothing to do with the mimetic: at M05 and M07 the seats' new flag is dominated by material in the English frame ("in her hands", "bath was ready") that was there under S204 too but that the seats no longer read past; at N07 the seats flag an omitted clause instead of an added mimetic sense.

Two readings of the null are consistent with the data and this design cannot separate them, but both withdraw §7.19's instruction. Reading (a): the deletion flip was largely an answer to the mutilation, and removing the phonaestheme without mutilating leaves nothing for a reader to catch — so the test S204 recommended is measuring the deletion, not the absence of the property. Reading (b): the plain substitute still carries the licensing property in another form (specificity, force, mouth-size, thinness, speed, aspect — the residues differ site by site — see RS-20260820c §7 and critic MAJOR 3), so nothing was subtracted, and the test's failure means the test does not remove what it claims to remove. Either way, "cover the mimetic and ask" is the wrong test.

Instruction (JA→EN, 擬音語・擬態語) — narrowed, in place of §7.19. Where the plain Japanese phrasing of the event is genuinely more general than the mimetic — as at N06, 音を立てて against つるつる・ちゅうちゅう, which drops the specific slurping-and-sucking the mimetic names — an English phonaestheme can read as a real addition, and the translator can hear it in the reading. The practitioner's test is not to cover the mimetic; it is to write the plainest phrasing of the event that carries no phonaesthetic form, and ask whether your English still says more than that plain phrasing. On this evidence that fires at 1 of 7 sites, not 7 of 9, so the licence the earlier instruction claimed — that any English phonaestheme at a Japanese mimetic is faithful to the mimetic and not to the event — is not supported.

The excluded sites are the data, not a technicality. M08 (the 「やに」 residue) and M12 (the 「ぱちつかせて」 residue) are exactly the two sites where a plain substitute could not be built without recreating a mutilation. Everywhere the substitute is clean, the flip has disappeared.

What is NOT evidenced.

1. Three named model seats, on seven curated items from one novel. Charter §4 forbids treating panel agreement as validation. Tier D NOT PASSED.
2. The mutilation-vs-residue readings are not separable by this design; separating them would need a per-item feature decomposition adjudicated by independent Japanese raters (critic MAJOR 3), which is not reachable in-session.
3. The mimetic-present baseline is not stable across prompt framings — this run's continuity miss at three sites is the evidence. Any subtractive test on this baseline inherits that instability.
4. Seven sites; substitutes preserve heterogeneous residues; the mechanical mimetic-form gate is naive; the same seats read the S204 items. Full statement in RS-20260820c §8.

internal-judgment-only. Full limits in RS-20260820c §8.

7.22 §7.18's zero survives a second coding hand — and the class the published precedent never answers turns out to be answerable at every locus, for no extra words — 2026-08-21 (S208), RS-20260821-matched-shape

ARM-published-figure step 2, and the arm closes resolved at 2 of 2. §7.18 told a practitioner that the one published English hand from the Arabic answers 0 of 18 matched-shape figures — muwāzana — and its own limits said the weakest joint was that all 72 of the census's three-way calls were made by the party holding the hypothesis. This section discharges that, and adds the thing the census could not say: whether the class can be carried at all.

The check. Three model seats, each given the whole 1819 chapter, one Arabic locus at a time, a flat content-only gloss, and a menu of formal relations fixed before dispatch — with no hypothesis, no counts, no translator's name and no knowledge that the anchor exists. The seat does not decide whether a figure is answered; it reports which relation it sees, and the strict and loose readings are computed from that. 162 bodies, 161 parsed.

Three things go into this handbook.

1. §7.18's zero stands, and the correction it took runs against this project, not for it. The lead's own coding of the third chapter had credited the published hand with a matched shape at two loci; three seats overturned both, unanimously, and the printed text supports the seats. At the five numbered virtues, the first … the second … the third … is a repeated frame, not a shared ‑th suffix — English ordinals do not rhyme with each other. At creation and production … death and destruction, the seats aligned to the three English words that actually render the three matched Arabic members and found no match on those. An independent hand made the published figure stronger by reading the correspondence more exactly than the party who wanted it. Cite the zero; cite §3.5 of the anchor with it.
2. The class is not inexpressible in English. It is unattempted. Impossible to express the sententious brevity of the Arabic with strict fidelity, wrote Knatchbull in 1818, and the measurement says he did not express it. But the same chapter rendered under a rule that forbids a silent reduction — R39: at every matched-shape locus, write MATCHED or write IMPOSSIBLE and say why, with the English quoted — comes back 24 MATCHED, 0 IMPOSSIBLE, sixteen of them matched by English morphology, at 2,348 English words on 1,284 Arabic against 1,849 per thousand for the same hand on the previous chapter without the clause. Forcing the answers did not lengthen the English.
3. The instruction, and it is a procedure rather than a preference. Where your source matches its members in shape and your target has no morphology to match them with, do not decide locus-by-locus whether English "affords" it — that decision is where the class quietly goes to zero. Write down, before you translate, what you will accept as a matched shape in English — matched suffix, matched inflection, matched word-class and length, matched frame — and then at every such place record either the match you wrote or the reason you could not. The list costs nothing to make and it is what turns a silent reduction into a visible refusal. Four of this hand's twenty-four matches were free (the plain English already matched) and are marked as such; a list that does not separate the free ones will flatter itself.

Limits. One translator, one work, one recension, three chapters, 1819 English. The seats are models, not human coders, the relation menu is unpiloted, and four of §7.18's eighteen loci remain unchecked. The 24 MATCHED is one hand's, under resources that hand declared in advance and called generous, and one of the twenty-four has already had its label corrected by the seats. The weakest rung of the menu — same part of speech, same length — fired at 8 loci and should not be reported as a match again. Tier D NOT PASSED; no jury was involved and nothing here is a quality claim about any rendering. internal-judgment-only. Full limits in RS-20260821-matched-shape §8.

7.23 §7.22's instruction survives; the check it invites does not — a matched shape has no plain version to be compared against — 2026-08-21 (S209), RS-20260821b-matched-heard

ARM-matched-shape-heard step 1. §7.22 tells a practitioner to declare in advance what will count as a matched shape in English, and then at every such place to record either the match he wrote or the reason he could not — and, under the regimes that implement it, the plain wording he refused at that span. That last record is what makes the instruction auditable, and it is what this section is about: it does not record a plain wording.

The measurement. Twenty-two loci with intact material, across two source languages and two renderings — the sixth Kalīla wa-Dimna chapter from the Arabic and 劉基《賣柑者言》 whole from the Chinese, the second rendered this session under R40 at twelve loci, twelve MATCHED, zero IMPOSSIBLE. At each locus the refused wording had been written at the same sitting as the rendering, before any experiment existed. Two blind seats were then asked the two questions a subtractive control needs to survive.

The primary was withheld before the run was dispatched. What the run then showed is that the screens describe the passages: three seats reading whole passages, prompted to list places where stretches echo in form, recovered the built matches at 0.879 of cells — and at the sixteen loci where the refusal still echoes they recovered a figure at 0.938 in the matched version against 0.930 in the plain one.

Two things go into this handbook.

1. Do not test a matched shape by writing the passage plainly. At a locus where the source matches two or more members in form, an English rendering that keeps both members and the same content matches them too — by frame if not by suffix — because the members stay in matched position and English supplies a frame there whether or not one is wanted. You cannot write the unmatched version, and a hand who thinks he has written one has, two times in three, written another figure: the places of destruction and the places of loss; guarding against what he fears and warding against what he dislikes; imposing enough to inspire fear, and dazzling enough to be taken as a model. This is the third subtractive check to fail in a month, and the first to fail because the property cannot be removed, rather than because the replacement smuggled a device in (§7.20) or the deletion mutilated the source (§7.21).
2. Know which half of the effect you are buying. Putting the members in matched position is the cheap half, and it is most of what a prompted reader registers — the flattened versions are heard as echoes at 16 of 22. Matching their suffixes or their measure is the expensive half, and it is what §7.22's declared list is really a list of. Both halves are worth doing; they are not the same work, and only the second one is optional. A corollary for reading §7.18 and §7.22: their zero is a zero for form-matching, not for parallelism, and it should be cited that way.

Limits. One hand, two works, three model seats, and no fourth reading seat available to buy. Each screen rests on a single seat, and the parity seat's calls are not all obvious (swallow against eat was judged DIFFERENT); what is robust is the direction and the size, not the exact set. The estimand is prompted detection — every seat was told to look for formal recurrence — so nothing here says an unprompted reader notices. All twelve Chinese matches used one resource, so no resource comparison is made or licensed. Tier D NOT PASSED; no jury was involved and nothing here is a quality claim about any rendering. internal-judgment-only. Full limits in RS-20260821b-matched-heard §6.

7.24 The ordinary English word does not tell you whether a sound figure can cross — the synonym does, and the members are not fixed either — 2026-08-21 (S210), RS-20260821c-echo-availability

ARM-alf-layla step 7. Three spans of The Thousand Nights and a Night had established that Arabic rhymed prose rarely reaches English, and §7's evidence base carried an unregistered guess about which loci cross: the ones where the plain English words for the two rhyming things already chime. The guess was registered before the fourth night was translated and it is false.

Fifteen rhymed loci, enumerated from the Arabic before any English for them existed. For each, two blind seats gave the ordinary English of each member one member at a time, with the siblings masked, so no seat could see a pair and rhyme on purpose. Three further seats then judged, blind to source and hand, whether any two expressions in a list echoed.

The registered tolerance was ≤ 0.10 and the direction is worth nothing (one-sided exact P = 0.452). Knowing whether the ordinary words chime tells a practitioner essentially nothing about whether a translator produced an echo.

Two things go into this handbook, and the second is the one to use.

1. The members are not fixed. RATIONALE WITHDRAWN 2026-08-23 (S216) — see §7.29: the rhyme is not lying in the English you have already written (0.000 gain at 50 seat-loci), and the clause "not lexical, it is structural" is refuted by the hand that executed it. At the copy-text's dawn formula — three cola on -āḥ, whose ordinary English the seats gave as {morning, appeared, handsome people} and {morning, appeared, sailors}, unanimously heard as no echo — Burton rhymes four deep, by writing a colon the Arabic has not got: when broke the dawn · and appeared the morn · and light was again born · and the Sun greeted the Good whose beauties the world adorn. A hand willing to add a member can extend a run. §7.18's and §7.22's counts of what "crosses" assume a fixed set of members and should be read that way.
2. Look for the synonym, not for the ordinary word. The proverb من لم ينظر في العواقب فما الدهر له بصاحب is carried by Burton as Whoso regardeth not the end, hath not Fortune to friend and, independently and without him open, by this project's own hand as He who does not look to the ends, Time is none of his friends. WITHDRAWN 2026-08-22 (S211) — see §7.25: the instruction below is refuted on a second language family and the withdrawal criterion fired. It looks like the paradigm case of a rhyme lying ready in the lexicon. It is not: asked for the ordinary English of العواقب, both seats said consequences. The chime had to be chosen — ends over consequences — and two translators a century and a half apart chose it because it chimed. The practical instruction: at a locus where the source rhymes, do not ask whether the first English word chimes; enumerate the ordinary synonyms for each member and look for a pair that does. The available rhymes live there.

Limits. One night of one work, Arabic→English only, three model seats and no human reader; the instrument counts a near-echo, and under a strict recount that admits only full rhymes every figure above falls without changing the verdict. The lead's own row is excluded from the primary on two grounds. Tier D NOT PASSED; no jury was involved and nothing here is a quality claim about any rendering. internal-judgment-only. Full limits in RS-20260821c-echo-availability §7.

7.25 §7.24 item 2 is WITHDRAWN — the rhymes are not in the synonyms, and the near-chimes are in the plain places too — 2026-08-22 (S211), RS-20260822-synonym-reach

ARM-synonym-reach step 1. §7.24 item 2, written yesterday off a single Arabic proverb, told a practitioner that at a rhymed locus the available rhymes live among the ordinary synonyms. Measured on a second language family, they do not.

The measurement. Sa'di's «گلستان» دیباچه — the densest rhymed prose in Persian, the project's first work in that language — with its saj' loci enumerated from the Persian and drawn by a salted hash before a word of English for the span existed. Two blind seats gave, for each of 76 rhyme-bearing words, one at a time with its siblings masked, first the single most ordinary English rendering and then up to six. Every cross-member pair was judged mechanically, by a pronunciation dictionary under a rule fixed in advance, not by a model.

Two full rhymes in 2,173 cross-member pairs. The registered yield bar was +0.25 and the registered withdrawal criterion was +0.10; the yield is +0.042.

Three things go into this handbook, and the first is a subtraction.

1. §7.24 item 2 is withdrawn as an instruction. Do not tell a translator that enumerating the ordinary synonyms at a rhymed locus will turn up a chime. On the class the instruction was written for, with up to seven candidate words per member from two independent hands, it turned up one, in twenty-four. §7.24's own paradigm — Burton's and this project's ends / friends for العواقب — remains true and is now visibly the exception it always was.
2. The warning that replaces it, and it is the strongest thing the run establishes. Relax the rule to admit a near-chime and reach goes from 1 in 24 to about 7 in 10 — and it goes to exactly 7 in 10 at the places the author left plain. The pairs that produce that number are spread / tend, put on / set, creatures / entities, scattered / suspended: nobody hears these as echoes. A translator enumerating synonyms and allowing himself a near-chime is not finding the author's figure, he is finding English, and §7.14 and §7.20 already say what to call that. AMENDED 2026-08-22 (S212) — see §7.26. The sentence "nobody hears these as echoes" was written on the lead's ear alone and is withdrawn as a general claim: a slant at a phrase-end is registered at +0.350 above a matched floor. It survives about its own examples — scattered / suspended, creatures / entities and spread / tend buy 0.000 — and the availability finding above, which is what the withdrawal of §7.24 item 2 rests on, is untouched.
3. §7.24 item 1 survives and is now the operative half of §7.24. SUPERSEDED 2026-08-23 (S216) — see §7.29. The move that actually produces an English echo at a saj' locus is not lexical, it is structural: change what the members are. Burton got a four-deep run at the dawn formula by writing a colon the Arabic has not got, and the one AFFIX chime the lead's own hand took here — whatever does not abide is not worth being tied to — rhymes on a word rendering a different Persian word from the one the source rhymes on. That is a real move with a real cost, and it is what the handbook should point at.

One thing worth keeping, against the withdrawal. The single STEM hit is ability / civility, for قوّت / مروّت — and it is the same pair the lead's own frozen log had taken at that locus, independently and before this design existed, where the first and most ordinary words are strength and courtesy. The procedure is not empty; it is rare. And the lead's hand found four chimes on ten STEM loci against the blind seats' one on twenty-four, which measures the hand, not the yield: three of its four used words no seat offered as ordinary.

Limits. One work, one span, Persian→English only, n = 12 loci in the primary class — the run establishes the near-zero and could not have detected a small positive. Two model seats, no human reader, no fourth reading seat to buy. The estimand is General American rhyme availability as CMUdict assigns it; some pairs are rhymes in other Englishes. The seats recognise the text (6 of 14 named it) and recall is not excluded, though it can only inflate a reach reported at 1 in 24. The lead's own figures exclude three loci where a published rendering was seen before the span was translated. Tier D NOT PASSED; no jury was involved and nothing here is a quality claim about any rendering. internal-judgment-only. Full limits in RS-20260822-synonym-reach §9.

7.27 The tradition's answer to rhymed prose is to drop it — four published hands, 128 chances, not one full rhyme — and the move into verse is not the answer either — 2026-08-22 (S213), RS-20260822c-persian-hands

§7.26 said, on reading evidence, that a chime is registered at +0.542 at a verse line-end and at +0.222 at a prose colon-end — that these are two moves and not one. This section says what translators have actually done about that, and one of the two obvious inferences from §7.26 is now refuted.

The record: «گلستان» باب اول ۱–۵, four published English hands over ninety-three years — Gladwin 1806, Ross 1823, Eastwick 1852, Arnold 1899 — at the 32 rhymed-prose loci enumerated from the Persian before any English was opened. Anchor A-gulistan-hands; result and limits RS-20260822c-persian-hands; verifier 35 checks, 0 failures.

1. A prose chime is not in the published record. Pooled over 128 cells the rate is 0.094, and the number graded STRICT is zero. Every one of the twelve answers is a near-miss, and most are the English inflection rather than a choice — applauded / mortified, laid aside / deposited, riches / years; two hands reach the same pair independently. For a practitioner: dropping a saj' is what every published English hand of this book has done, at every one of these places. It is the tradition's practice, not a lapse.
2. Moving the passage into verse is NOT what they do instead. This was registered as P2 and it fails: the genre change is taken at 4 of 128, all four by Arnold, who prints most of the book in English verse and still leaves 28 of his 32 saj' loci in prose. The inference from §7.26 — that hands spend the chime where the payoff is — does not hold.
3. What the same hands do in verse is the contrast. At the 42 bayts, Eastwick answers 0.905 and Arnold 0.952, against 0.094 and 0.031 in prose in the same volumes: registered as P3, gaps +0.811 and +0.921 against a ≥ 0.30 bar. Gladwin and Ross answer 0 of 42, because they print no verse at all: the first half-century of this book in English carries none of its sound.
4. A caution that costs nothing to state and would have cost the finding. At S11, where Sa'di rhymes two verbs, three hands and the lead all produce an audible English chime — men / women — and none of them is answering the figure: those words render the nouns. A chime in the neighbourhood of a source figure is not carriage, and a study that grades clause-ends rather than the words that render the source's rhyme-bearers will count it as carriage four times out of four.
5. Availability is not the explanation. Under R43, which forces the attempt and forbids the escape into verse, one hand reached a prose chime at 20 of 32 loci, on renderings an outside check found to add nothing (8 of 8 planted additions caught). The pairs are there; the published hands were not looking for them. That hand is contaminated and is not evidence about English in general — RS-20260822c §7 states the limit.

Evidenced on: Persian→English, one author, one span, four hands. Not evidenced on any other pair or author, and it makes no causal claim that the verse-setting responds to the sound figure — the saj' loci are aphorisms and the controls are narration, note (bqr).

Extended and bounded by §7.31 (2026-08-24): item 1 holds on a second author in a second language; item 3 does not generalise — two of three English hands of al-Ḥarīrī print the verse as verse and leave it unrhymed; and item 2's GENRE-CHANGED measure cannot see a hand who lineates the prose without rhyming it, which is what Preston 1850 does, so any "hands move rhymed prose into verse at rate X" figure is unsafe until that definition is settled.

7.28 Run length: four is reachable in English, five ran out, the price is repetition rather than invention — and the tradition's answer to a four-deep run is two pairs — 2026-08-23 (S214), RS-20260823-run-depth

§7.26 and §7.27 are about whether a chime is carried. This section is the first thing this project has said about how far — the extent of a figure rather than its presence.

A Persian قطعه holds one rhyme across three or four consecutive bayt-ends and leaves the hemistichs between them bare. In English, one line to a hemistich, that is a rhyme at every second line-end. The record: «گلستان» باب اول ۶–۱۳ whole under the new R44, which requires the English rhyme to fall where the Persian rhyme falls and forbids converting a run into couplets; and the four published hands at the three runs of depth ≥ 3. Result and limits RS-20260823-run-depth; verifier 260 checks, 0 failures.

1. Depth four is reachable and costs nothing extra; depth five ran out. All 9 of 9 runs were held at the source's depth, 13 of the 14 links full rhymes by tools/rhyme_pairs.py. At the four-deep V09 the hand found four distinct bearers (farewell / well / fell / tell). At V11, whose مطلع makes five positions, the English rhyme family ran out and the run was held only by using two bearers twice — the fault Persian prosody calls ایطا. For a practitioner: the constraint is the size of the English rhyme family at the meaning you need, and it binds between four and five. Sa'di pays nothing because his rhyme sits on a grammatical ending, so his family is as large as the morphology; an English end-rhyme sits on a lexical ending, so its family is as large as the dictionary allows at that sense.
2. Holding the run does not buy content. Two seats reading the Persian, on an instrument that caught 6 of 6 planted three-word additions by name, flagged the forced arm at 5 of 9 passages and the tradition's couplet arm at 4 of 9 — and cleared V11, the hardest run in the span, in both seats. The price of depth is repetition, not invention. At four of the five flagged places the seats named the very words the translator's log, frozen before the check existed, had already recorded as bought by the rhyme.
3. No published hand has ever held one. 0 of 12 cells reach the source's depth, by the bearer measure and by the looser position measure alike. Six of the twelve are not verse at all — Gladwin 1806 and Ross 1823 print all three passages as running prose. Neither verse hand holds one rhyme more than two lines deep anywhere. What they do instead is specific and worth copying or refusing on purpose: Eastwick answers a four-deep run with abab cdcd — two pairs at the source's own distance, the rhyme changing at the halfway point, and the four bare lines rhymed as well. Arnold answers with couplets and enjambs across every bayt boundary, so no English line corresponds to a hemistich at all.
4. Availability is not the explanation, and one hand refutes it inside the passage. At V07 Eastwick writes a sextet rhyming ababab — two interleaved chains at distance 2, in the very passage where Sa'di runs one three deep — and not one of his rhyme-words renders one of Sa'di's. He could hold a rhyme at that distance and depth. He did. He did not do it there.
5. Nothing here says a reader hears the difference. The reading measurement built to ask it saturated — three seats found every rhyme in every condition, adjacent and at distance alike, 1.000 against a false-alarm floor of 0.003 — and was withheld by its own registered criterion. The preregistered contrast was not observed; that is not evidence that distance costs nothing. Note (brb).

Evidenced on: Persian→English, one author, one span, one hand forced by a regime, four published hands at three runs. Not evidenced on any other pair or author. It makes no claim about reception, and item 1's boundary between four and five is one hand's reach in one rhyme family.

7.29 §7.24 item 1 loses its rationale: the rhyme is not lying in your draft either — and what a hand actually does is BOTH lexical and structural — 2026-08-23 (S216), RS-20260823c-member-move

ARM-member-move step 1. §7.25 item 3, written 2026-08-22 in the act of withdrawing §7.24 item 2, promoted §7.24 item 1 to this handbook's only surviving positive instruction about carrying a source's sound figure:

The move that actually produces an English echo at a saj' locus is not lexical, it is structural: change what the members are.

It stood on two instances and no rate. A rate was registered and measured, and it is zero.

The measurement. Sa'di's «گلستان» باب اول ۱۴–۱۹, 27 prose blocks, with the whole enumerated population of saj' loci and plain parallel colon pairs frozen from the Persian before any English for the span existed. Two blind seats, each shown one colon at a time and never a sibling, gave a plain literal English rendering of the colon and the ordinary English of its rhyme-bearing word; chime decided mechanically. The manipulated variable is the size of the pool the chime may come from — the source's own rhyme-bearer alone, against every content word the plain rendering already puts on the page, which is the reservoir "change what the members are" assumes.

The registered withdrawal criterion was < +0.10 and the gain is 0.000. Pool-size-matched: 0.000. Dictionary-excluded: 0.000. Restricted to the four words nearest the clause end: 0.000.

Three things go into this handbook, and the first is a subtraction.

1. §7.24 item 1's availability rationale is WITHDRAWN by name. Do not tell a translator that the rhyme is somewhere in the English he has already written and needs only to be moved to the end of the colon. At fifty seat-loci of the class the instruction was written for, widening the search from the source's rhyme-bearer to the whole colon turned up nothing that was not already there. This is the same bar, the same instrument and the same work that withdrew item 2 the day before at +0.042.
2. "Not lexical, it is structural" is WRONG, and both halves are now measured near zero separately. The lead's own hand, under a regime that forced the move and priced it (R45, log frozen before dispatch), took 12 chimes at 29 rhymed loci — and every one is built on English words that appear in neither blind seat's plain rendering of the colon. The hand did not find a rhyme in its draft; it rewrote the draft until one was there, choosing a different word and moving which word carries the rhyme, in one act. All 12 are RESELECT; REORDER — put a different word of your own colon last — was used zero times in 42 attempts. So the operation the instruction names is not the operation that works.
3. The instruction that replaces both is about price, not about search. At those 29 loci the hand carried 12, refused 10 that the same instrument scores as full rhymes, and found nothing at 5. hateful/ungrateful for دون, thief/grief for خیانت, peace/cease for فراموش کردند: each is a real English rhyme and each says something the Persian does not. About as much of this craft is in declining the available chime as in finding it, and a translator who does not keep both counts cannot tell which he is doing. A regime without an AVAILABLE-REFUSED category would have reported 26 chimes here and been wrong by ten.

One thing carried forward, not new. Near-chime at the widened pool is available at 0.420 at rhymed loci and 0.556 at the cola Sa'di left plain — the same direction as RS-20260822-synonym-reach's 0.708 against 0.708, on a fresh span and a different pool. §7.14 and §7.20 acquire a third instance: relax to a near-chime and you are finding English, not the author's figure.

Limits. One work, one span, Persian→English only; STEM primary 25 loci, control 10 loci and declared underpowered before dispatch; two model seats and no human reader; the estimand is CMUdict's General American. W2 is a ceiling on availability in a plain rendering and not a measure of what a move can reach — a blind realisability check was priced and refused in writing. Tier D NOT PASSED; no jury was involved and nothing here is a quality claim about any rendering. internal-judgment-only. Full limits in RS-20260823c-member-move §8.

8. What v0.2 is NOT evidenced for

Everything v0.1 §4 lists, unchanged, and one addition:

• S1 is a statement about what five runs failed to find, not a proof that the sites do not exist. Five pairs, five instruments, one project, no human readers. A rate this project cannot detect is not a rate of zero, and §2's «Что-с?» is a live counter-instance.
• Trajectory-level marking is measured in ONE pair only (§6, S140). Russian→English carries a vocative-borne change and not a pronoun-borne one, at status released-on-repair; the second pair returned no licensed primary. v0.2 is not evidenced for any claim about trajectory marking outside Russian→English, and the one-pair claim rests on a repaired gate rather than a blind one.
• Nothing about register carriage, which is Q-e, and the refusal has now been re-reasoned three times (§7, §7.1, §7.2). §7 states a problem; it recommends nothing, and the four published hands it rests on are all 1890–1918, so a period norm and a translator's decorum are not separated. S145 added that the two devices are separable and that the cheaper one is not placeless (§7.1) — it did NOT add a recommendation, and the register comparison between them is withheld for want of power. S150 removed the want of power — 20 non-tied sites against a bar of 12 — and the comparison was withheld anyway, because the located-idiom permission was exercised at 4 of 60 hand-sites (§7.2). S155 built the from-source instrument §7.2 asked for and withheld the primary a FOURTH time, on an instrument control: one of the two judges cannot see a located idiom that a frozen translator's log documents. What S155 adds is not a recommendation but a defect in everything above it — the placeless baseline these differences are measured from is itself placed by both judges, on items including a national SPELLING, so §7's whole comparison has been run against a floor nobody measured (§7.3).**
• S1 generalises over five pairs that all have English as the TARGET, and the reversed direction does not reverse it — it replaces it. With English as source, R1's loss does not arise at all (v0.1 §7's rewritten EN→ row): 0 sites of 26 and 0 of 7 in two censuses. S1 is therefore not confirmed, not refuted, and not applicable in this direction, and the thing that happens instead has no entry in v0.2 §2: an obligatory target category the source does not fill. Nine of nine published-and-lead hands fill it symmetrically where the English marks nothing and asymmetrically where the English marks rank — a description of translator behaviour, not a finding about loss. v0.2 is not evidenced for any claim that the symmetric filling costs anything, and two runs built to measure that have now been stopped by their own registered gates (RS-20260809d §4, RS-20260810 §4).
• Nothing about which CARRIER a source mark uses, and the first design built to find out was stopped by its own source gate. RS-20260811f's control had shown a Bengali social relation reaching English at 0.8333 when carried by an address noun and at 0.0000 when carried by a verb ending, which is a conjecture with a predictor in it: a distinction the source expresses in a free word reaches English; the same distinction in a bound morpheme does not. RS-20260811g crossed carrier with content — social and epistemic — inside one Turkish story, and G1, the source-parity gate, failed: three arbiters reading the Turkish agreed the variants differed at only 0.4444 (bound) and 0.5556 (free) against a registered 0.667, and dropping the one seat that plainly could not read Turkish still leaves the bound arm at 0.5833. Every primary is withheld, and the withheld difference (+0.3556, 90% interval [0.189, 0.544]) would have failed its own 0.40 bar anyway. What the run does establish is only what its two passing controls establish: a bound morpheme English must carry crossed at 1.0000, and adding a semantically inert free word crossed at 0.0000 — so neither bound marks are invisible nor any added word looks different can explain a carrier effect if one is ever measured. v0.2 is not evidenced for any claim about carriers, and S1's sentence "where the source's marking rides on a device English also has … the relation transfers" still rests on the five pairs of §2 and on no test of the device itself.
• Nothing about the SIZE of the realia effect. §7.6 is at ceiling on all eight passages and §7.9 is at ceiling on every rung of the ladder; a forced pairwise choice at 1.000 is a floor on detectability, not a magnitude. v0.2 is not evidenced for any claim that a realia decision changes an absolute verdict about the English — §7.5 looked for exactly that and found +0.048 at n = 7. ~~And nothing about moderate domestication: a translator who domesticates four sites in forty is not described.~~ PARTLY DISCHARGED S167, §7.9 — moderate domestication is now described on the prose channel, item by item, on ten windows in three source languages. It is still not described on the world channel, where §7.9 measured both doses against the untouched page and neither against the other.
• Q-e's ELEVATION half: the instrument's resolution was measured before the question was, and the two results point opposite ways. RS-20260814d (S183) built a three-rung ladder on Danish and its resolution gate failed at +0.0833 against +0.25; it deposited the reading that a one-step policy does not raise and only prevents a fall. RS-20260815 (S188) refutes that on Russian→English at +0.7500, and refutes the rival span-length explanation too at +1.0000 (§7.11). What v0.2 is therefore not evidenced for: any general statement about what a moderate register policy does, in either direction. Two pairs, two authors, opposite results, and the most economical account of the pair — that the Danish arm's rule had not fired at the sites in question, since both arms there read Engaged! at one of them — is post hoc and unmeasured. And nothing at all about where published translators sit above plain English, which is Q-e's original question: §7.11's arms are all one lead's, and RS-20260814d's published-hand quantities were withheld by its own gate and remain withheld.
• Nothing about quality, still. Tier D is NOT PASSED.

10. Where a source's footing mark lands in English — the section ARM-footing was constituted to write

Written 2026-08-13 (S176) by ARM-footing step 2. Rewritten whole 2026-08-14 (S186) by ARM-honorific-hands step 2, and that arm closes here. The rewrite is not a patch: the section carried a struck-through sentence, a correction appended after it, and a §10.5 whose question had been answered. What follows is the section as the evidence now stands, with the superseded wording gone rather than crossed out. The argument for each change is in the results cited, and the strikethrough version is in this file's history at 6e65f67.

Four language pairs: RU→EN, BN→EN, DE→EN, JA→EN. Five sources of evidence: RS-20260812i-footing-channel (Russian), RS-20260812h-dakghar-grade (Bengali), RS-20260806e (German), RS-20260813e-slot-typology-ja and RS-20260814b-honorific-hands (Japanese), plus the R06/R27 pair of §10.5.

Read this first: most of this section is a refusal, and the one new thing in it is a price list. ARM-footing was constituted to give S1 a predictor — to say where a translator should expect a grammatically-marked social relation to land in English. It found one, tested it at the first opportunity, and it did not survive. ARM-honorific-hands then removed the section's one positive sentence, on the materials §10.5 asked for. What is left standing is stated first, then what is not.

7.26 A slant IS registered — but not the three slants §7.25 named, and mostly only in verse — 2026-08-22 (S212), RS-20260822b-echo-threshold

ARM-echo-threshold step 1, and it corrects a sentence this handbook published the same day. §7.25 item 2 told a practitioner that nobody hears a near-chime. That was the lead's own ear with nothing behind it, and it is now measured.

The measurement. 25 constructed loci from the Gulistan دیباچه — 14 prose, 11 verse, in the project's own English for both — each built as a 2 × 3 crossover so that the phonetic relation is present in exactly one of four cells and every main effect of either word cancels in the interaction. Levels assigned mechanically by tools/rhyme_pairs.py. Three seats, 414 detection calls, 0 dead. Pre-run critic two rounds, 4 BLOCKING; the crossover exists because of one of them.

Five instructions, and the first is a correction.

1. "Nobody hears these" is withdrawn as a general claim. A slant at a phrase-end is registered, at about six-tenths of what a full rhyme is worth. If you have written one, it is not nothing.
2. It was right about its own three examples, and that is the useful part. scattered / suspended, creatures / entities, spread / tend buy exactly nothing — +0.000, −0.333, +0.333 across three loci. Some slants carry and some do not, and the enumeration procedure §7.25 withdrew is precisely a machine for producing the ones that do not. The distinction, not a blanket permission, is what replaces the blanket prohibition.
3. Position decides most of the value. A slant at a verse line-end is worth +0.542, most of a full rhyme; at a prose colon-end, +0.222, about a third. Answering saj' in prose with a slant buys a third of what answering an end-rhymed couplet with one buys. These have been treated as one move in this handbook and they are two.
4. What you buy is a half-chime and is named as one. Of 37 slant detections carrying a strength verdict, 35 were called half; of 46 full-rhyme detections, 1 was. No reader is going to mistake your slant for a rhyme. Price it accordingly.
5. Say registered, not heard. The instruments read text. On a dedicated twelve-item probe they called non-rhymes that are spelled alike — sword / word, beard / heard, cough / dough — an echo at 0.722, against 0.778 for real rhymes spelled unlike; one of the three was doing spelling and not sound at all. A chime that works on the page may not survive being read aloud, and this handbook cannot yet tell you which of yours is which.

Against the finding, and reported because it is against it. The cell carrying the slant was ranked the least natural English of the four by an independent naturalness pass — so detection happens despite the wording reading worst, not because of it. And the fluency screen failed its own control (3 of 4 planted breaks) by rejecting the Gulistan's elevated register as "not ordinary English", so the item set's fluency is unverified.

Warrant, on the face of the instruction: 20 constructed loci from one work in one language pair, three model seats, no human reader, Tier D not passed. Unprompted, even a full rhyme is spontaneously remarked on only 0.389 of the time against 0.800 when the seat is told to look; everything above is what a reader attending to sound recovers.

7.30 The half-chime reaches the reader through the eye, and only the reader who is not listening — 2026-08-24 (S217), RS-20260824-eye-or-ear

ARM-echo-threshold step 2, and the arm closes here. §7.26 item 5 left a practitioner unable to tell whether the chime he had written works on the ear or on the page. This section puts numbers under that, withdraws a piece of §7.26's generality, and gives §7 back one small positive instruction — the first since §7.29 emptied the shelf.

What was measured. 14 constructed prose loci, each a pair of parallel clauses, in a 2 × 4 crossover. Two fillers carry the same rule-class sound relation to the phrase-end word — a pararhyme, NEAR-N2, same coda and a different vowel — and differ only in whether the pair also looks alike at the rime: full / gull against full / coal. A third filler is a full rhyme spelled unlike; a fourth is nothing. Levels assigned mechanically by tools/rhyme_pairs.py. Three seats, 536 calls, 0 dead. Pre-run critic two rounds, 23 findings, and this run went ahead under a NEEDS-REDESIGN verdict — see item 5.

Four instructions and a warning.

1. A pararhyme at a prose colon-end buys nothing. coal against full, seed against flood, boat against height — the consonance you reach for when the rhyme will not come — was registered 0 times in 28 by a reader attending to sound, against a full rhyme at 0.964. This narrows §7.26 item 1: that section measured the whole NEAR class at +0.222 at a prose colon-end, and the N2 sub-class is carrying none of it.
2. If you write a half-chime, write the one that also looks alike at the end of the word. That is the positive instruction, and it is small: full / gull, not full / coal. It is worth +0.286 to a reader who is simply reading, and about nothing to one who is listening.
3. Say seen, not heard, of a half-chime. §7.26 item 5 hedged registered against heard and could not say which. For the N2 class the answer is now on the page: the effect is there when the reader is not attending to sound and gone when they are. A translator's half-chime is working on the eye. (This says nothing about a full rhyme, which was registered at 0.964 under the sound-framed prompt and is unaffected.)
4. The question §7.26 item 5 asked cannot quite be posed, because English barely supplies the case. Of 50 classic spelled-alike English pairs, 35 are pararhymes, 4 are stress mismatches, 1 is an actual rhyme, and 10 come back with no sound relation — of which six are r-coloured pairs (word/sword, heard/beard, earth/hearth) that all end /rd/ or /rθ/ and are misfiled by the rule. What is genuinely alike in spelling and unrelated in sound is the -ough family and little else: 4 of 50. An English eye rhyme is a consonance four times in five. Two of the three pairs §7.26 item 5 prints by name carry a real sound relation, so its probe was never the clean orthographic contrast it was read as.
5. The warning, and it is not small. The second critic round returned NEEDS-REDESIGN and its surviving BLOCKING finding is that the two fillers are different words, so what is measured is an association — spelled-alike pairs in this class are registered more — and not a decomposition into orthography and residual sound-distance. The naturalness control did not fire but came within 0.07 of firing, in the confounding direction: the spelled-alike cell reads more naturally by 0.43 of a rank. Items 2 and 3 should be read as the spelled-alike half-chime is the one that gets noticed, with the reason undetermined.

Warrant, on the face of the instruction: 14 constructed prose loci in one language pair, two reporting model seats and one control seat, no human reader, Tier D NOT PASSED, and a NEEDS-REDESIGN verdict standing against the design. Stage d was also dispatched twice by an operator error, so the reported figures are the first call per cell and a declared sensitivity over the repeats is printed beside them (RS-20260824 §11a); no instruction here turns on which is used. The same accident measured that the seats are not deterministic at temperature 0 — 28 of 368 repeated cells changed their answer. The seats were chosen on prior results (RS-20260822b stage P), which is outcome-conditioned selection, declared before dispatch. Nothing was measured at a verse line-end, where §7.26 found the chime worth two and a half times what it is worth in prose; every instruction above is a prose-colon instruction. The interval is a dispersion statistic over 14 hand-built loci, not an inference about English or about readers.

7.31 The record holds outside Sa'di — three English Maqāmāt, 1767–1867, one full rhyme in 187 chances, and two translators who write down the policy — 2026-08-24 (S218), RS-20260824b-hariri-hands

§7.27 told a practitioner that dropping a saj' is the tradition's practice. Its evidence was four English Gulistans: one author, one language, four hands who read each other. This section is the same census on a second author in a second language — al-Ḥarīrī's first Maqāma, the work that defines Arabic rhymed prose — and it adds the thing a count cannot supply: two of the three translators say in print what they are doing and why.

The record: «المقامة الصنعانية» whole, 590 tokens, 65 primary loci frozen from the Arabic before any English was opened; Chappelow 1767, Preston 1850, Chenery 1867. Anchor A-hariri-hands; result and limits RS-20260824b-hariri-hands; verifier 335 checks, 0 failures; pre-run critic one round, 14 findings, 5 BLOCKING, all accepted.

1. §7.27 item 1 survives the move to a second author and a second language. Per hand the ANSWERED rate is 0.070 · 0.108 · 0.108, each at or barely above that hand's own accident rate in the same book (0.057 · 0.071 · 0.048); excluding the suffix rule, 0.035 · 0.046 · 0.062. STRICT is 1 in 187 cells — not the zero §7.27 recorded, and the honest gloss is that one is what accident predicts: Chenery's own background STRICT rate is 0.0048 over 210 pairs, which over 187 cells predicts about one. For a practitioner: on seven published hands, two authors, two languages and 133 years, the answer to a source rhyme in prose is to let it go. That is now the tradition's practice in two traditions.
2. And two of the three hands say so themselves. Chappelow 1767, in the note on the first figure of the book: "nor shall I imitate the author in my translation. To attempt it might be looked upon as a piece of pedantry: and indeed our English tongue will not admit of it." Preston 1850, in his introduction: "Rhyming prose is extremely ungraceful in English, and introduces an air of flippancy, unless the subject be of the most light and frivolous description." The behaviour and the declaration agree, and the one hand with no such statement found is the one that produces the study's single full rhyme.
3. What English keeps is the doublet. COLA-KEPT 190 of 195 = 0.974 — Preston and Chenery lose not one of the 65 paired members; Chappelow, who expands and rewrites, loses five. This was registered in advance as the study's one positive prediction and it is the practitioner half of the finding: saj' has two parts, and the parallel members cross into English prose intact while the sound does not. It confirms, on a third work and with a definition fixed beforehand, what RS-20260814 §1 observed on Lane and Burton.
4. There is a fourth move on the board, and no one uses it. §7.27 named three answers — drop, chime in prose, lift into verse. Preston takes a fourth: he sets the prose out in lines, one Arabic colon to a line, evenly balanced and unrhymed, and names it in his preface as "a species of composition which occupies a middle place between prose and verse". It answers the shape of saj' typographically and pays nothing in sense. No English translator this project has read since 1850 does it.
5. Printing the verse as verse is not rhyming it, and §7.27 item 3 does not generalise. All three hands print al-Ḥarīrī's two poems as verse. Preston answers 9 of 9 — four-deep best / zest / test / rest, then six aab stanzas chained across pairs — and Chappelow and Chenery answer 0 of 9. On the Gulistan every hand that printed verse answered at about 0.9; here two of three do not.
6. A measure of §7.27's that cannot see Preston, stated rather than repaired. GENRE-CHANGED is read off the page — indented, line-broken, capitalised at the line-head. On that rule Preston's prose is verse at 65 of 65; on his own printed declaration it is prose at 0 of 65. The same hand, 1.000 or 0.000, on a definition. Any figure of the form "hands move rhymed prose into verse at rate X" is unsafe until that is settled, and this study reports both rather than picking.

Evidenced on: Arabic→English and Persian→English, two authors, seven hands, 1767–1899. Not evidenced on any other pair. No causal claim about verse-setting responding to the sound figure: al-Ḥarīrī's unrhymed prose in this maqāma is six words long, so no matched control exists and the comparison is against a background accident rate, which is weaker — note (bqr), firing a third time on a third language. ANSWERED here is a pronunciation dictionary's verdict, not a reader's; §7.30 measures what a reader registers and is a different object.

7.32 The chime is registered at both placements, and moving it one line further off costs about a third — on two seats of three — 2026-08-24 (S219), RS-20260824c-run-placement

§7.28 said run depth is reachable and claimed nothing about whether anyone registers it, because the task built to ask saturated (note (brb)). This is the rebuild, and it is the first thing §7 has said about a device's EXTENT rather than its presence.

Seven four-line units of the lead's own English rendering «گلستان» باب اول ۲۶–۳۱, each in five placements of a single chime, each pair differing in one word, judged by which reads better — a question that never mentions rhyme, sound or line-endings.

What a practitioner is told, and it is narrow. A قطعه's rhyme placement is not inert. Carrying one rhyme across a run of bayt-ends, as R44 requires and as no published hand has ever done (§7.28, 0 of 12), buys about two thirds of what the same chime buys at a couplet's end — not nothing, and not as much. A translator who converts a run to couplets is not discarding an effect nobody could register; a translator who holds the run is not writing for nobody.

Three things bound it, and the third is the largest.

1. The construct is preference under explicit comparison, not hearing. Two passages side by side differing in one word make that word salient. Nothing here says a reader hears a rhyme.
2. The gap between the placements lives entirely in their control cells. hit(D1) and hit(D2) are identical (0.714 each); the +0.107 is the difference between the two controls. Referenced to the untouched passage instead, the extent effect is 0.000. Take registered at both placements and leave the gap alone.
3. Two seats of three produce all of it. P2 returns +0.571 / +0.357 and QR +0.429 / +0.500; P1, the frontier reporting seat, returns +0.071 at both placements — near nothing. The pooled figure is P2's halved. And QR's largest figure is at the Persian's distance, which removes the reading in which distance-2 is intrinsically hard to register.

Evidenced on: Persian→English, one author, one span, one hand, seven loci, three model seats. Not evidenced on any other pair, and on no human reader. Tier D is NOT PASSED. D1 is a distance control, not the tradition's couplet form, which supplies twice the chime; nothing here compares the Persian's placement with what any published hand does.

And a craft finding from the translation limb, frozen before the design existed. The span's four-deep قطعه was held with four distinct full-rhyme bearers and no repetition — the first clean one of the three this project has met. The reason sharpens §7.28's: English does have a grammatical rhyme family, the weak preterite in ‑ed, and when the sense wants past-tense verbs at the rhyming positions the family is as large as the verb list. The two runs that could not be carried are both runs where Sa'di's rhyme sits before a radif, and ~~English has no slot before a repeated word to rhyme in~~ — CORRECTED 2026-08-26 (S224), §7.36. English has that slot whenever the radif is something an English clause can end on. Both of the failures this sentence generalised from were the transitive-verb case, which is the one case where it does not. Sa'di's ghazal ۴۹ was carried at 8 rhyming positions of 8 with the chime in the pre-radif slot.

7.33 The oldest reason in the record for not carrying a sound figure is measurable, and on this pair it is right about the cost — 2026-08-25 (S221), RS-20260825b-flippancy

§7.31 recorded that two of three published English Maqāmāt print a refusal to carry al-Ḥarīrī's rhymed prose, and that both give the same kind of reason — a reason about the reader. This section is what happened when the object those reasons are about was built and put to a blind panel.

Preston, 1850 — "Rhyming prose is extremely ungraceful in English, and introduces an air of flippancy, unless the subject be of the most light and frivolous description." Chappelow, 1767 — "To attempt it might be looked upon as a piece of pedantry."

T-maqamat-sanaa-R48-v1 renders the whole first Assembly with a chime at the colon end wherever the sense allows — 33 full English rhymes against the restrained R43-v1's 2 — and T-maqamat-sanaa-R48D-v1 is that text with one colon-end word changed per chime and nothing else. Fifteen segments, three arms, three seats, blind to the whole question.

All three of the declared reader-effects are registered, in the direction the two translators named, on within-seat standardised scores over 15 paired segments:

The instruction that follows, and it is narrow. A translator who decides to carry a source's prose rhyme into English is buying a lighter, less graceful, more self-displaying English, and this is the first measurement of that price rather than the fourth assertion of it. Whether the price is worth paying §7 does not say, and does not have the evidence to say. What it can now say is that the reason the tradition gives for refusing is not folklore: three seats that were never told rhyme was at issue reproduced all three of its terms.

What §7 may NOT take from this.

1. Not that rhyme causes it. The two arms differ in the chime and in the diction the chime forced, and 41 of the 65 de-chiming substitutions make the control the more faithful text. The estimand is a rhyme-forward policy, which is the bundle a translator actually chooses — not the sound in isolation.
2. Not Preston's exemption, which is withheld. Its gate asked whether the panel reads the two strata as differing in subject; the sermon came back grave 8 of 8, and the comic half of al-Ḥarīrī's most comic maqāma came back light only 2 of 7. Preston's condition — "the most light and frivolous description" — is not instantiated by the picaresque frame of a maqāma, so his exemption could not be tested on his own author. RS-20260825b §3.
3. Not as a reader measurement. Three model seats, Tier D NOT PASSED, provisional. Preston's readers were Victorian and human, and this result is the strongest argument the project has yet produced for building independent human readers.
4. One work, one hand, one language pair, one dose, one maqāma of fifty.

And one finding that cuts across the instruction. The largest single effect in the run is at a locus where the rhyme was not reached for: Come near and dine; or, if you would rather, stand up and opine is one of the four full rhymes the restrained R43 rendering found by taking the ordinary word for both bearers. The panel scored it 8 · 8 · 7 on levity with the rhyme and 1 · 0 · 0 without it. The price §7.33 names is charged on any chime that lands, whether or not a policy went looking for it** — which makes it a fact about English at that locus rather than about a regime.

7.34 The price §7.33 named is not decisive, and what makes it not decisive is telling the reader what the source does — 2026-08-25 (S222), RS-20260825c-worth-paying

§7.33 ended on a sentence naming its own gap: "Whether the price is worth paying §7 does not say, and does not have the evidence to say." This section is what happened when the question was put as a choice rather than as a set of ratings — and the first thing it found is that the question cannot be asked without context at all.

Three real policies on one text, all by one hand, all frozen with logs before any design existed: restraint (R43, 2 full colon-end rhymes), the rhyme first (R48, 33), and — new this session — R50, Preston 1850's own printed remedy, the positive half of his sentence that nobody had ever instantiated: no rhyme at all, the members of each saj' pair balanced in syllable count and shape, no colon over fourteen words. Three blind seats, both presentation orders, three information conditions.

1. The registered primary is WITHHELD by its own position gate, and the gate is the finding. Told nothing at all, the seats chose the passage shown first at 0.819; on a control pair of one text against itself with a single word changed, they did the same at 0.739. Split by order: shown the rhymed rendering first they took it 29 to 7, shown it second they took it 6 to 30. A forced choice between two competent translations of the same passage, offered with no context, is barely more content-driven than a forced choice between one translation and itself.

2. What survives is a statistic position bias cannot produce — a segment counts only when the same arm wins in both orders. It is post-hoc, invented after the gate failed, and is reported as such.

3. The instruction, at the size the evidence supports. A translator's decision to carry a source's prose rhyme is not defensible on the page alone and is defensible once the reader knows what the source does. The seats do not stop noticing the cost when they change their minds — they name it and pay it: "Its playful rhyme better reflects the original's rhymed prose, despite slightly more contrived diction." The specificity control holds: where neither arm rhymes, the identical disclosure moves one segment each way, so this is not a general wish to please whoever supplied the fact.

4. Preston's remedy is the best-liked text when nothing is disclosed, on its first instantiation in 176 years — but he described a practice and never predicted a preference, so this is a modern consequence of a Preston-like prescription and not a verdict on his claim; and both its figures are 0.0625 tails that do not survive the run's own Holm correction.

What §7 may NOT take from this.

1. Not that a reader prefers anything. Three model seats, Tier D NOT PASSED, every figure provisional and internal-judgment-only. The claim at issue is about reception and this is not a reception instrument — the second section running to end at that sentence.
2. Not a registered result. The primary is withheld and the headline is post-hoc. A successor that registers the both-orders statistic in advance is owed before this is cited as settled.
3. Not an effect of rhyme. The estimand is preference among these three particular translations, which differ in the policy and in everything the policy dragged with it.
4. One seat supplies most of the size — 22 to 2 against 15 to 9 and 14 to 10 — though all three lean the same way and none reverses.
5. One hand, one work, one maqāma of fifty, one pair, and the same segments §7.33 used.

And one craft finding from the translation limb, which needs no seat at all. Refusing the chime as a policy is not the same act as merely not reaching for it: an ordinary careful rendering of saj' (R43) fell into 29 near-echoes and 5 identical colon-end words by accident; the rendering that refuses them has 5 and 0. Four of the eight refusals R50 had to make are the same accident — al-Ḥarīrī's second-person sermon ends colon after colon on the enclitic ‑كَ, whose ordinary English is a final you, and English cannot rhyme you with you. The balanced period also pushes the diction toward commoner words: R50 has 4 colon-end bearers outside CMUdict against R43's 15 and R48's 18.

7.35 Preston keeps half of his own rule: the length is real and it is fourteen words; the balance leaves no trace — 2026-08-26 (S223), RS-20260826-balanced-period

§7.31 and §7.33 quote Preston 1850's sentence; neither had checked his page. He refused the rhyme and printed a positive remedy in the same breath — clauses "though not rhyming together, arranged as far as possible in evenly balanced periods, and never exceed a certain length." His own English of two Assemblies was measured against it, beside Chenery 1867, who declared nothing, and beside R50, the lead executing the sentence as a rule. The two halves come apart.

The length holds, and it is recoverable. Across 230 printed lines of «الصنعانية» and «الحلوانية» not one exceeds fourteen words (maxima 14 and 13; CV of length 0.198 and 0.177; syllable range 6–19 and 7–20). The line is a real unit, not a scan artefact: 95% of the recovered lines end in punctuation and 99% begin with a capital. And it is imposed, not inherited — the Arabic cola he renders have CV 0.475 and 0.433, two to three times his own, and he prints 107 and 123 lines for 139 and 140 cola, merging the short ones.

The balance does not hold. On the registered primary — mean adjacent-length difference divided by the mean over 10,000 reorderings of that text's own lengths, so a hand with uniformly short units earns nothing for it — Preston is smoother than Chenery in one cell of six (three segmentations × two Assemblies), and on «الحلوانية» under his own lineation his ratio is 1.005, p = 0.55: exactly what a shuffle of his own lines would give. R50, the same sentence executed deliberately, is the smoothest text in its panel under a segmentation applied identically to every hand (0.653 and 0.731), so the quantity is reachable; Preston has not reached it.

Everything he controls, he controls at the line and nothing above it. Under a rule that splits only at strong punctuation his 95th-percentile unit is 54 and 61 syllables against Chenery's 42 and 46 — his short even clauses chain into long uneven sentences.

What the handbook may now say. The one positive prescription in the English record for what to put where the rhyme was is a length discipline, not a balance discipline — and the length, on the page of the man who printed the prescription, is about fourteen words to the clause. What it may not say is that balanced periods are unreachable, or that Preston's readers noticed either half: this is a measurement of two pages, with one declarer and no replication of the declaring, and RS-20260825c is the standing record that the project's preference instrument could not say whether any of it is worth having.

One coincidence, recorded as a coincidence. R50 rule 5 declared a fourteen-word ceiling the day before, from the Arabic's median colon, by a hand that had never opened Preston's translation. His observed maximum is fourteen.

A finding that is not about balance at all. Scored on adjacent line-ends, Preston has 6 IDENTICAL and 7 NEAR in 106 pairs on «الصنعانية» — five of the six are consecutive thee in the sermon, which is likely deliberate epistrophe, but the seven near-rimes (delusion ⁄ oppression, neighbour ⁄ observer ⁄ retainer ⁄ ruler) fall at exactly the places al-Ḥarīrī rhymes, in the book whose introduction calls rhyming prose ungraceful. §7.34's craft finding — refusing the chime is not the same act as not reaching for it — now has a published hand in it: R50 on a fresh maqāma has 0, 0 and 0 in 139 pairs, and the man who printed the refusal has 6 and 7.

7.36 Whether a Persian radif can be carried into English is decided by the radif's part of speech, and it is decidable before the first line — 2026-08-26 (S224), RS-20260826b-radif

This section is a craft finding and it corrects §7.32. It rests on six ghazals rendered whole and on no model judgment at all; the run that was bought is in §7.36.4 and established nothing.

A Persian rhyming position is […qāfiya][radif]: a word or phrase repeated identically at every rhyming position, with the rhyme on the word immediately before it, and nothing between them.

7.36.1 — the rule, and it is grammatical. A Persian radif can be any word, because Persian puts the verb last and a Persian clause therefore ends wherever the poet likes. An English clause ends on an object, a complement or an adverbial, so an English radif can only be something an English clause can end on — and the qāfiya slot immediately before it inherits that constraint.

Written as a prediction table before any of the six was attempted, and 6 of 6:

So a translator can know, before writing a line, whether the shape is available — and §7.32's generalisation was drawn from two Gulistan failures that were both the transitive case.

7.36.2 — a radif repeats a form and lets the sense move under it; English repetition drags the sense along. Sa'di's برخاست means rose up at six positions of ۵۰ and departed at two. English arose covers the first six and there is no third option, because a radif is one word by definition. The hand picks one sense and pays at the others, and which sense to buy is a real decision the form forces and the source does not record.

7.36.3 — at a comparative radif English supplies a chime whether or not one is wanted. Every English comparative ends in the same unstressed syllable, so a rendering that must put a comparative at every rhyming position cannot be given a no-chime control: the attempt graded NEAR at 3 of 3 pairs and the locus was withdrawn. This is RS-20260821b-matched-heard's obstacle from the other side — there English supplied a matched frame at matched positions unbidden; here it supplies a chime. Two devices, two runs, the same wall in front of the subtraction method.

7.36.4 — and the measurement that was supposed to price the repetition half failed, for a reason worth printing. Four line-end arms per window (chime × repetition, 2 × 2), seven windows, three blind seats, both orders, 168 forced choices, 0 dead. All three registered predictions are unestablished — two withheld by the design's own order criterion, one failing outright at a median of exactly 0.000. The diagnostic: the seats returned the same answer when the two passages were swapped at 54.8% of cells, against a chance rate of 50%, and all three preferred whichever passage was shown first (0.625 to 0.768). On minimal pairs differing in three line-end words, a forced-choice preference task reads position at least as strongly as it reads text. §7.32's published figure is not impeached — it pooled both orders, and a symmetric position effect attenuates a pooled difference rather than biasing it — but any future preference instrument here is measured for order-consistency before it is believed.

Evidenced on: Persian→English, one author, six poems, one hand, and — for 7.36.4 — seven constructed windows and three model seats. Tier D is NOT PASSED. 7.36.1–7.36.3 are craft findings from a frozen translator's log and carry no model judgment; 7.36.4 is internal-judgment-only and provisional.

7.37 A translator's printed REFUSAL of a device predicts what a reader loses; his printed PROMISE to keep one does not survive the same test — 2026-08-27 (S226), RS-20260827-declared-play

ARM-declared-function step 2, and the arm closes resolved at 2 of 2. Everything §7 tells a practitioner rests on a premise nobody had checked: that a translator knows what his own choices do to a reader. The arm tested it on printed declarations rather than on labels this project manufactures, and it now has both halves.

What holds. A translator's printed account of what a device does to his reader predicts what a reader gets — when the account is a refusal. Three refusals in the record, three confirmations: Preston 1850 on flippancy (§7.33, measured and confirmed on his own author), Chappelow 1767 on pedantry (the largest of the three effects), Knatchbull 1818 on the sententious brevity of the Arabic (A-knatchbull-kalila §3.1, 0 of 18 matched shapes answered, exactly as he said). A refusal is also the cheap half: a man who prints that he will not do a thing will be found not doing it.

What does not hold, and this is the new evidence. Eastwick 1852 printed the opposite kind of declaration — "I have uniformly done my best to preserve the play upon words which occurs so often, and which is accounted such a beauty in the East" — and it is the first declaration this project has tested that could be caught failing. At 22 تجنیس loci in 14 prose blocks of «گلستان» باب دوم, enumerated from the Persian by script with every candidate's disposition published before any English was opened, three blind seats reading one passage at a time found a play in Eastwick's English at 0 of 14 blocks — the same as Gladwin 1806, who promised nothing, and below Ross 1823, who also promised nothing and scores 1. The seats returned slightly more plays in control passages whose Persian has none. Q1 is withheld by its own registered gate, which fails in that direction.

The instrument is not blind, and the bound matters more than the null. On the identical passages and loci, the same seats detect the lead's R52 rendering — the same span rendered with the play chased at every locus — at 10 of 14, reference tail 0.000977. So a play of the size a translator who is trying produces is visible to these readers; the three published hands do not produce one.

Two qualifications the framework must carry with the sentence.

1. The promise is not empty. In a block the study's own rule had classified as a control, Eastwick writes "deprived of their society … derived profit from this story", and two of three seats found it — answering وحید / مستفید at exactly the locus. The rule was blind to that pair. The honest negative is not "he broke his promise" but "at the places this study could enumerate, blind readers found nothing of it."
2. Two of the three hands are DEPENDENT? — EAS~GLA share a 12-token run, and Eastwick's own footnote on this span says "Gladwin and Ross translate as above, and I am content to follow them." A-gulistan-hands may not be pooled as four independent witnesses.

What a practitioner can take from §7.37, and it is a warning and not a technique. Your own account of what you are protecting is worth most where it is an account of what you are giving up. A stated refusal, with a reason about the reader, has three times been found to describe a real effect. A stated intention to preserve has been tested once and left no trace a blind reader could find — including in the pages of the man who stated it, who was still right about one locus in a place nobody was looking. Count what you kept; do not trust the sentence in your preface.

Status. provisional; three model seats, no sense Tier-D calibrated. Evidenced on one pair (Persian→English), one book, one chapter, ten tales, plus the three earlier refusals on Arabic→English and Persian→English. The alignment key was built by one hand that knew which translator was which, which RS-20260827-declared-play §10.2 records as the strongest surviving objection. Limits in RS-20260827-declared-play §10.

7.38 Telling a reader what the original does moves their preference; showing them the original does not — 2026-08-27 (S227), RS-20260827b-shown-or-told

ARM-worth-paying step 2, and the arm closes resolved at 2 of 2. §7.33 ended on "Whether the price is worth paying §7 does not say, and does not have the evidence to say." §7.34 answered it — the price is not decisive once the reader is told what the source does — on a post-hoc statistic invented after its own gate failed, and named registering that statistic in advance as the first thing owed. This is that run, on a maqāma the hypotheses were not formed on, in two renderings that did not exist when they were.

It confirms the effect, refuses the mechanism §7.34 implied, and the refusal is the useful half.

What holds. A sentence telling the reader that the Arabic is rhymed prose moves preference toward the rendering that carries the rhyme. Within-cell, holding segment, seat and presentation order constant: 7 of 54 cells move toward the rhyme-forward arm and none moves away — exact one-sided 0.0078, Holm-adjusted 0.0156, and the segment-clustered check agrees at 6–0. All three seats lean the same way. The flips are a seat reversing its own stated reason on text it read one call earlier: "forced rhymes that make the phrasing feel unnatural" becomes "brilliantly recreates the distinctive rhymed prose (saj') … delightfully poetic."

What does not hold, and it is new. Showing the reader the original does not do what telling them about it does. Given the same passage's Arabic — in the script and in a machine romanisation, with no sentence describing it — preference does not move at all (6 up, 5 down). The direct told-versus- shown contrast separates them: 1 cell up, 7 down, two-sided 0.070, clustered 0–5. And the seats say why: shown the source they stop judging English and start checking words — "more accurate word choices ('kings' for aqyāl, 'getting a living' for iktisāb)".

The control that sharpens it. A de-chimed copy of the rhymed arm — the same rendering with the sound removed at 49 colons and, by script, nothing else — behaves differently under both conditions: shown the source, readers move toward the plain arm against it (2 up, 10 down, two-sided 0.039); told, it does not move (5 up, 4 down) where the chiming arm moves 7–0. The registered three-part specificity conjunction fails at its first clause because the shown primary is null, and the direct between-block difference on the told contrast reaches only 0.108. Suggestive, not established, and the page says so.

What a practitioner can take from §7.38. If you carry a source's formal device, say so in a note; do not rely on the reader inferring the warrant from the original. On this evidence the paratext does the work the facing page does not. A reader given the original starts checking your words, and a marked English word order — which is what a prose chime costs, forty-two times in the rendering measured here — is exactly what that check penalises.

What §7 may NOT take from this.

1. Not that a reader prefers anything. Three model seats, Tier D NOT PASSED, every figure provisional and internal-judgment-only. Third section running to end at this sentence.
2. Not that the source's form is inert. The shown condition is longer, more salient, and invites source-checking; the design declared in advance that it measures a total effect and does not decompose it. Showing did not work here is not the original cannot do this work.
3. Not a large effect. In 46 of 54 cells the seat chose the same passage whether or not it was told, and one seat supplies 4 of the 7 flips.
4. §7.34's own headline does not replicate as a statistic. The both-orders content-decided count, which carried §7.34 at 7–0, was registered in advance here and reads 3–2; the pre-run critic showed it cannot reach significance at nine segments whatever it shows. §7.34's finding survives in the told direction; its statistic should not be cited again.
5. Preston's balanced period is absent from this run, because it and the restrained arm measured as near-duplicates on this text (§7.34's second finding is untouched, not confirmed).
6. The control for the contrast that moved was bought after the primaries were seen and is flagged unregistered throughout.

Status. provisional; three model seats, no sense Tier-D calibrated. Evidenced on one pair (Arabic→English), one work, one maqāma of fifty, nine segments, one hand. Verifier 461 checks, 0 failures, 3 of 3 mutations caught; pre-run critic two seats, both NEEDS-REDESIGN, and the design was rebuilt — the inferential unit moved and a control arm was built — before any data call. Limits in RS-20260827b-shown-or-told §9.

10.1 What is evidenced: the loss, on four families and five strata

S1-a. Where a source grades the speaker against the addressee in a slot English fills compulsorily and with one form — above all the personal pronoun — the marking does not reach English, and translators do not put it anywhere else either. Russian ты/вы: 0 of 11 (RS-20260812i). Japanese 貴公 / お前 / 手前 / 己 / わし across thirteen sites and four hands: 0.022 (RS-20260813e). Bengali: a total loss at every grammar-only site in both hands (RS-20260812h). German: King 1914 carries no device at 9 of 9 (RS-20260806e).

Two further strata, never measured in this project before, were measured on two published human hands in 2026-08 and are lost at the floor (RS-20260814b, 『源氏物語』 ch. 15 「蓬生」, Suematsu Kenchō 1882/1900 and Arthur Waley 1926, 67 mechanically derived sites, two blind cross-assigned coders):

• addressee-polite はべり — the Japanese 丁寧 stratum, which grades the hearer and not the referent: 0.000 over 19 rendered sites;
• the honorific prefix 御 — a courtesy on a noun: 0.000 over 14 rendered sites, against a registered prediction of ≥ 0.35, which fails.

The sharpest single case the project holds is Japanese 手前, which in one Akutagawa story is a contemptuous you to an equal and a self-abasing I to a daimyō. All five hands render it you at one end and I at the other. None shows that it is one word.

The practitioner instruction that follows, and it is narrower than the one this section carried until 2026-08-14. At these sites, no hand this project has observed carries the mark, and a translator who believes the relation is being carried should check — five pairs now say it is not. What S1 already said — expect most sites to need nothing — is unchanged, and S1-a says why for one class of them. What the instruction may no longer say is that nothing can be recovered. Every hand behind the numbers above was translating without being asked to carry the grading, and a census of hands that were not trying cannot tell English has no device from translators decline to spend one. §10.5 separates those two, and the answer is the second.

10.2 What is NOT evidenced: the predictor, withdrawn before it was ever recommended

RS-20260812i §3.1 proposed, post hoc, that carriage is predicted by a property of English — the Russian deference clitic -с reaches English at 42 of 43 because it sits at the utterance end and English's utterance-final position is free, so sir drops into it. E-20260813e registered that as a prediction and tested it on Japanese deferential predicate endings, which sit in the same place.

It failed. Carriage 0.200 against a registered bar of 0.50. And the mechanism failed harder than the number: across 1,762 words of hypothesis-blind direct speech — Glenn Shaw's published 1930 English and the lead's own English of the same story written before the account existed, plus two 2026 machine hands — the utterance-final vocative is used to carry a footing mark once, and that once is sarcastic. Shaw uses no sir, no my lord, no your lordship at all.

A fourth candidate failed the following day, and it died before it was ever written down here. RS-20260814b's G5 predicted that 御 would cross because it attaches to a noun and English has nouns. It crossed zero times of fourteen: 御ありさま becomes your circumstances, 御心 becomes his disposition. The prefix is a courtesy, and an English possessive is not a courtesy, so the slot English puts the word in has nothing to put the courtesy in. A source mark riding on a part of speech English also has is not thereby carried.

No positional predictor and no part-of-speech predictor enters S1. ARM-carrier failed to find one twice and closed retired with the refusal written (§2); ARM-footing is the third failure and the first that had a stated mechanism to kill rather than an instrument gate to trip over; RS-20260814b's G5 is the fourth. A framework that recommended "look for the free slot and re-spend the mark there" would have been recommending something no translator in this project's evidence base does.

10.3 What actually carries in a published human hand, and it is a verb

Across 5,361 words of English in five hands on 『源氏物語』 ch. 15 — two published human translators, two 2026 language models given a prompt that never mentions honorifics, and one hypothesis-aware lead rendering excluded from every number — every site carried by consensus of two blind coders falls into exactly two classes.

Class 1, and it is not news: the source also supplied a noun. 故宮 the late Prince, 大将殿 His Excellency the Commander, 式部卿宮の御女. Wherever a hand carries anything, it carries it here, and it carries because the Japanese handed it a title to translate. That is S1 as already written.

Class 2, and it is one clause in 5,361 words. Waley, at the three sites where the aunt's deference to the Princess is purely grammatical — 思し隔てて and the double honorifics 渡らせたまはね / 許させたまへ:

…but even if you will not deign to have any dealings with us yourself, I am sure you will not be so inconsiderate as to stand in this poor creature's way…

The English device that carries a Japanese grammatical honorific in a published human hand is a deferential auxiliary verb. No other hand in the run carries a grammar-only site at all. A post-hoc mechanical count over the whole periphrasis-and-rank vocabulary (deign, vouchsafe, be pleased to, condescend, humbly, humble, exalted, presume, venture to, beg leave) — registered nowhere, offered as description only — finds WALEY 3, P3 1, SUEM 1, P1 0, lead 0.

What this section said until 2026-08-14, and why it is gone. It said that where English carries this at all it does so by putting a rank-noun where the pronoun would go (the pipe that is now in Your Lordship's hand), and refused to recommend it on the ground that "the only hands that do it are the two 2026 language models." Both halves are withdrawn. Given a second Japanese source, the machine hands use the device zero times, and across all five hands the whole rank-noun set occurs once, in Waley, as a vocative ("But, Madam…"). What the earlier section had observed was not a machine habit and not an English resource: it was that 「煙管」 contains a retainer addressing an actual daimyō, so a rank-noun was available to be used at all, and this chapter's quarrel between an aunt and a princess supplies none.

deign is not recommended by being counted. One occurrence, one hand, one clause; Tier D is NOT PASSED, and nothing here says Waley's English is better than Suematsu's, who writes no such thing.

10.4 The insertion claim, withdrawn on two failures to replicate

RS-20260812h §5 found both English hands of Tagore's «ডাকঘর» inserting a contemptuous epithet where the Bengali is neutral, a century apart, and proposed a reallocation of the marking to the translator's discretion. That claim is withdrawn. It failed to replicate in Russian (0 insertions, 40 unmarked paragraphs, 7 hand-story cells) and again in Japanese (0 insertions, five hands, both coders, ten mechanically derived negative-control paragraphs). v0.2 does not say that translators habitually re-spend the marks they lose. On the evidence the Bengali insertions are local to one scene.

10.5 What a translator who tries can reach, and what it costs — untested

§10.5 used to ask a materials question — find two published human hands and census them — and that question is discharged: RS-20260814b found them, censused them, and its numbers are in §10.1 and §10.3. What replaces it is the question the census could not answer, because every hand in it, and every hand in §10.1 and §10.2, was translating without being asked to carry the grading.

ARM-honorific-hands step 2 therefore produced the missing arm of the comparison: 『源氏物語』 ch. 15 §3-4, 1,227 characters, rendered twice by the same hand in the same session — once under R06 (source-only, single pass, no instruction about footing) and once under R27 footing-max, a regime that instructs an attempt at every site and requires the price of each attempt to be written down. 52 sites, extracted mechanically and committed before either rendering; the baseline frozen before the site rows were read. T-genji-yomogiu-R06-v2, T-genji-yomogiu-R27-v1, workshop/regimes/R27-footing-max.md.

Everything in this subsection is untested in the sense of D-20260724-04: the reachability figures are one hypothesis-aware hand, self-reporting on its own sentences. It may be cited for what was reachable at what price. It may not be cited for what English does, for which the hands that were not trying remain the only evidence, and it is not a recommendation to translate this way — the R27 rendering is deliberately worse English than its own baseline.

What is no longer self-report, as of 2026-08-14: the price. RS-20260814h-footing-price put a third rendering between the two ends — T-genji-yomogiu-R28-v1 under R28 footing-selective, 12 of the 52 sites carried inside a declared budget, +3.5% words, 12 marked tokens — and a bulk-matched decoy (the baseline plus 177 words of ornament carrying no deference), and put all four to three blind seats. The seats are models, not readers, and no sentence below says otherwise.

1. Nothing was unreachable. 52 sites: 51 carried, 1 abandoned, 0 where no device came to mind. So the sentence this section carried until today — there is nothing to recover and no craft that recovers it — is wrong in its second clause, and §10.1 no longer contains it. The loss measured on five pairs is a loss translators take, not a loss English imposes.

2. The price, and it is the practitioner-relevant number. 15 of the 51 carried sites cost nothing; 36 cost something named. The rendering runs 836 → 1,012 English words on the same 1,227 characters, +21.1%, and carries 78 marked deference tokens in 1,012 words, one every 13.0 words (corrected 2026-08-14 from the 74 first published: the counting script matched on the file's raw text and four devices straddled a line break), against zero in the baseline. The Japanese marks arrive one every 23.6 characters. The rate of marking is comparable; the visibility is not. Japanese footing marks are inflections on words that had to be there anyway; the English ones are 176 extra words. A translator who carries all of it has not reproduced a feature of the source — it has added one, which is the §7.8 diagnostic (a mark the reader can see that the source's reader could not) arriving at §10 from the other direction.

2a. And the price is now decomposed, which changes what a practitioner should be told (RS-20260814h, 72 blind bodies, verifier 66 checks 0 failures). The maximal rendering loses 1.389 points of naturalness against its baseline. A decoy that adds the same 177 words of ornament, and no deference at all, loses 0.972 of them. So only 0.417 points — 30% of the drop — is attributable to the deference devices, and that figure missed the 0.50 margin the design registered in advance. Seventy per cent of what carrying the footing costs is what adding 21% of any words costs. The instruction that follows is not this is expensive but this is mostly bulk: the first thing to buy is fewer words, not different ones.

2b. The decoy also answers the CLUNKY worry in the reassuring direction, on this item. RS-20260809h had a word-scramble scoring +1.4583 above the prose it scrambled on perceived-source-carriage. Here the ornament decoy scores 2.389 on the social-marking item against the plain baseline's 3.111 — 177 words of conspicuous padding lowered perceived social marking rather than raising it. One item, one comparison; it does not rehabilitate perceived-source-carriage, which is a different item.

3. A fifth candidate mechanism, and the first that predicts which sites rather than whether. It is post hoc, was noticed while coding, and is registered here as something to test, not as a result. Every English deference device the attempt found is volitional — deign, be pleased to, vouchsafe, condescend, see fit — and Japanese たまふ is not. たまふ attaches to whatever a superior does, suffers, feels, or simply is. So the devices fit where the source's predicate is something its subject chose, and misfit everywhere else. Sorted that way, the 31 subject-honorific sites split 9 of 20 free among volitional predicates against 0 of 11 among non-volitional ones, and the single abandoned site is non-volitional. The misfit is not stylistic: his lordship was pleased to be sorry for her turns an involuntary pang into a granted favour, and your ladyship has been pleased to pass hidden away in these weeds says the Princess chose the ruin the whole chapter exists to say she did not choose.

4. One stratum had no device of the right kind at all. Addressee-polite はべり grades the hearer and says nothing; the only English found was an explicit comment on the speech act (— I say it with all deference —), which is a different category of thing and which no reader can un-notice. The alternative — lifting the whole clause's formality — is not localisable to a site, and RS-20260814g has just measured that readers hear such a lift as ornament the source does not have. This is the stratum §10.1 records at 0.000 over 19 sites in two published hands, and it is the one place where the census's silence and the attempt's silence agree.

5. Court verse carries no footing marks, and the two renderings are word-for-word identical across both poems. Nothing follows from it for S1; it is recorded because it is the only part of the passage where the regimes could not differ.

6. NEW 2026-08-14, and it is the first thing §10 can say about what the marking delivers. The devices are perceived, and what they deliver is a relative grading whose direction is not the source's. Blind, the maximal rendering is scored socially marked in 18 of 18 cells against the baseline's 11 of 18; among its marked answers the modal reading is one person above another (12) rather than both above the narrator (6). But the direction follows the local density of devices: in the paragraph where her ladyship occurs ten times all three seats name the woman as the higher, and in the paragraph the narrator's his lordship was pleased to dominates the same three seats answer both / the man / both. A translator carrying every mark is not carrying the source's grading; it is manufacturing a grading whose direction is set by which character each paragraph happens to be about. That is a cost §10.5's price list could not see, and it is not paid in words.

6a. CORRECTED AND STRENGTHENED 2026-08-15, RS-20260815c-footing-direction. Item 6's stated basis was too strong; its claim survives on better evidence. The WHO item item 6 rests on had never been shown to separate the wording from what the paragraph narrates, and a matched manufactured pair now shows it does not: on a scene that narrates a woman being deferred to, adding four rank words moves REL by +2.33 and moves the named person not at all; on the same scene with the deferential behaviour removed, the same four words take REL from 1.00 to 6.33 and the answer from nobody to a named person, 3 seats of 3. Narrated behaviour dominates wording on the direction of a perceived grading; wording survives on its intensity. §10.5 item 6, §10.5 item 7 and RS-20260814h §6–§7 must all be read with that beside them. The direction claim itself is now on evidence where the events are controlled by construction. T-genji-yomogiu-R06-v1 and T-genji-yomogiu-R33-v1 are the same text word for word but for 70 words of footing marking, and three blind seats part company on who is marked highest at 8 of the 10 segments — the baseline naming the lady of the house at 3 segments and the marked rendering at 9. Four renderings of the same 1,841 characters — Suematsu 1882, Waley 1926, and the lead's two — are read as putting different people on top at 9 of 10 segments, and 9 of 10 again among the three that render the span at length. Both figures are post hoc: they were registered after the run's own instrument gate failed and before the data existed, and every registered primary of that run is withheld.

6b. NEW 2026-08-15, and it is not about carriage at all: a title on an absent character outranks any amount of grammatical marking on a present one. At the two paragraphs where the aunt is pleading with the Princess, the Japanese marks the Princess against the aunt eleven and twelve times. In Suematsu and in the lead's unmarked baseline, three blind seats unanimously name Genji — who is not in the room and never speaks — because Prince Genji and His Excellency the Commander are the only rank nouns in the paragraph. A translator deciding what to do about a source's honorifics is not working on a blank page: the rank vocabulary the plot forces into the prose is already telling the reader who is highest, and it is louder.

7. NEW 2026-08-14, post hoc and offered as a candidate design, not a finding: half of this passage marks the social relation in English without any grammatical device at all. In three of the six segments the seats' social-marking scores are driven by nouns present in every arm including the unmarked baseline — the Princess, her women, the Dazai Deputy's wife — and the judges said so in as many words. Where the baseline names ranks, the devices add 2.111 points to a score already at 4.556; where it is silent, they add 4.556 to a floor of 1.667. §10 is written throughout as though the grammatical channel were the only one, and on this evidence the relation arrives by a lexical road in half the text, uninstructed. Nothing in §10.1's five-pair census measured that channel, because it was counting honorific morphology.

10.6 What is still open

• ~~Whether the price in §10.5 is real to a reader. That design does not exist yet.~~ BUILT AND RUN 2026-08-14, RS-20260814h-footing-price, and the answer is in §10.5 items 2a, 2b, 6 and 7. The pair was priced blind, with a third rendering between its ends and a bulk-matched decoy beside it. The price is real and it is mostly not the footing's. Three things it did not settle and which are the successors:
• Whether a selective purchase is audible is UNRESOLVED, and the run withheld it rather than reporting a null. The middle rendering carries 12 devices in 865 words — one every 72 — and scored +0.200 above the baseline, but the design's own noise control fired: on a segment where the two arms are the same text under different labels, one seat scored it 4 and then 1. A twelve-token purchase is not distinguishable from a label on this instrument, and saying so is not the same as saying the purchase is inaudible.
• ~~The direction problem of item 6 has no remedy in §10. If maximal carriage manufactures a grading that points wherever the paragraph points, then a rule for how much to carry is the wrong shape of rule; what is wanted is a rule about whose footing a paragraph is allowed to mark.~~ WRITTEN, EXECUTED AND ANSWERED IN THE NEGATIVE 2026-08-15, RS-20260815c-footing-direction. The rule exists: R33 footing-directional (workshop/regimes/R33-footing-directional.md) reads each paragraph's own marks, identifies the party they place highest, and requires that party to carry the most marked material and no other party to be marked — a dominance rule, obeyed at all twenty graded paragraphs of the censused span at +6.0% words. It does not deliver the source's grading. It delivers a smoother wrong one. The source reverses direction at three paragraphs — a princess using addressee-polite はべる to her aunt, an honorific verb on a dead nurse, servants elevating the gentlewoman who is leaving — and R33 solved all three in writing, with no rank word in any of the three solutions. None of the three reaches a blind judge: at two of them the seats answer the lady of the house, on the her ladyship in the paragraphs on either side. Obeying the rule made the rendering more uniform than any hand that ignored it, and the uniformity is wrong in exactly the three places the source is not uniform. The successor question is not a smaller rule of the same shape: it is whether anything below the scene is the right unit for a footing decision at all, given that English marks with lexis and lexis persists across paragraph boundaries where a Japanese inflection does not.
• The lexical channel of item 7 has never been measured and is not a honorific question at all.
• Whether the volitionality account survives a registered test. It was noticed while coding the thing it describes, which is the weakest possible provenance. It is testable cheaply: it predicts which sites a published hand carries, and two published hands are already censused site by site in RS-20260814b.
• Period, and it is unseparated everywhere in this section. Both published Japanese hands are pre-1930 (1882, 1926); RS-20260813g carries the same limitation for the same reason. Nothing here separates what English does from what English did before 1930.
• NEW 2026-08-15: nothing in §10 has ever controlled for what the passage narrates. Every perception figure in §10.5 and every one in RS-20260814h was measured on passages whose events differ along with their wording, and §10.5 item 6a now records what that is worth. A design that wants to attribute a perceived social grading to the translator's words has to hold the incident flat, and the only clean way this project has found is two renderings that are the same text apart from the marks.
• NEW 2026-08-15, and it is the first thing in §10 that runs the other way: RS-20260815d-supplied-footing puts English on the SOURCE side and finds that the RELATION crosses a source that marks it in no grammatical slot. Doyle narrates a footing reversal on a fixed dyad in forty lines of "Silver Blaze" and marks none of it grammatically; six Japanese renderings — 三上於菟吉 1930, its 2021 revision, the lead, and two blind 2026 hands — all mark Silas Brown as rising toward Holmes, by +1.333 to +2.000 points of a four-level 対者敬語 scale, agreeing on the direction at every site. So §10's measured loss is a loss of the grammatical channel and has never been shown to be a loss of the relation. What carries it is the utterance's own lexis and illocution plus the identity of the parties, not the narrated behaviour: a hand shown one line and nothing else reproduces 60–67% of the shift, a hand shown the line plus "spoken by Brown to Holmes" reproduces all of it, and the sentence in which Doyle states the reversal adds nothing on top of the two names. The obvious extrapolation from RS-20260815c — that the scene is where the information lives — did not happen here. "Your instructions will be done," a flat English passive with no politeness in it, is rendered in 敬語 by nine renderings of nine. The counterpart, and the practitioner-relevant half: three blind seats reading the English lines out of context answered cannot be determined from the words alone at two of six sites, and every Japanese hand marked those sites anyway and disagreed about them. ~~§10 owes a written subsection on the lexical channel and on the silent sites, and ARM-supplied-footing step 2 is where it gets written.~~ WRITTEN 2026-08-16 AS §10.7, and it is mostly a refusal: two independent readers of the same 70 Dickens utterances agree on whether the English fixes the footing at 0.471, so the subsection cannot carry the instruction it was wanted for. RS-20260816d-lexical-channel. Limits that bind it: one scene, one pair, one published human hand, six sites, and the two blind hands are measurably not independent of each other (33-character shared run — note (bpi)).
• Nothing in this section is about quality. Tier D is NOT PASSED.

10.7 The subsection §10.6 owed — and it cannot be written as an instruction, because "where the English says nothing" is not a determinate set — 2026-08-16 (S198), RS-20260816d-lexical-channel

ARM-supplied-footing step 2, and the arm closes here. §10.6 has carried, since 2026-08-15, a promise to write two things down: the lexical channel — where English marks footing at all, it marks it in the utterance's own lexis and illocution — and the silent sites, where the English says nothing and a translator into a marking language is inventing. The second was to license the one practitioner instruction this section has never been able to give: the places to watch are not the loud ones.

It cannot be given, and the reason is structural. Two independent readers of the same English, neither of which translated anything, were shown 70 utterances of Dickens one at a time, with no speaker, no addressee and no scene, and asked where the speaker stands relative to the person addressed — with cannot be determined from the words alone offered as a real answer in the same sentence as the scale.

1. They agree on how much footing a line carries and not on whether it carries any. On the nine sites where both gave a number they agree within one point on 9 of 9 of a seven-point scale. On whether a number was possible at all they agree at 0.485 — worse than chance on a binary. One reads X at 0.871 of sites, the other at 0.353. Half the census (35 of 70) is a site one reader calls determinate and the other calls silent.
2. The disagreement is principled, not incompetence. Shown "Both very busy, sir" the conservative reader answers X and writes "'Sir' is polite but alone does not establish a definite social hierarchy"; the other answers 1 and writes "the use of 'sir' signals deference … a social superior." Both have read the word. They differ on whether a politeness token is evidence about rank. The same split runs through the cordial sites: one treats 4 = equals as a claim needing evidence, the other as the default.
3. So the instruction §10.6 wanted has no denominator. Watch the places where the English says nothing presupposes that those places are identifiable. On this census they are identifiable to one reader in nine cases of seventy and to another in forty-four, and the two sets overlap on nine. A translator following the instruction would be told to watch a different half of the scene depending on who drew the map.

This is the second consecutive session to reach the same shape on a different phenomenon. RS-20260816c (§7.15, written earlier the same day) found that "the source's sound figures" is not a determinate set — the lead's inventory and three readers of the Arabic agreed on 13 of 21 loci. Two different framework instructions, on two different channels, in two different language pairs, both fail at the same place: the set of sites the instruction quantifies over cannot be fixed. §7.15 and this item should be read together; the successor question belongs to both.

What the gate did, and it is registered. CTL1 — the three sites carrying an explicit English sir must all be classed determinate — failed at one of three, exactly because the conservative reader answered X to "Both very busy, sir." By the frozen design that withholds the run's three primary predictions. They are withheld and not reported, and the scene-A translation cells that existed only to compute them were not bought.

What survives, and it is the craft half.

• Q5 HOLDS 3 of 3, and its SWAP control says what the marking is made of — this is the session's sharpest number and it is a warning, not a technique. Five Japanese hands — 森田草平 1929, the lead, and three blind 2026 hands — mark Mrs Cratchit's speech to her husband polite and his to her plain, at 7 of 7 sites, on English that grades neither and in a scene where she wins the argument. Re-dispatch the same seven English lines with the words husband and wife exchanged, and the gap flips sign in 3 of 3 hands, from +0.694 to −0.250. Bob's four words "My dear, Christmas Day." come back as 「おまえ、今日はクリスマス じゃないか」 labelled husband and 「あなた、今日はクリスマスですよ」 labelled wife. The deference is in the label, not in Dickens's sentence, and it travels with the label. What looked like two independent hands agreeing about the Cratchits' marriage is a shared convention keyed to two nouns. The practitioner consequence: a footing choice made at a silent site is not a reading of the source, and a reader of the translation will take it for one.
• The lexical channel is real but does not cross on the axis §10 measures. The one site in the whole scene-B census where the English carries an explicit politeness marker is the clerk's "If quite convenient, sir." 森田 renders the deference into an address term — 「ご都合が宜しければ、貴方。」— with no polite predicate at all, which the project's frozen 対者敬語 coder scores 0; the lead's 「お差支えございませんようでしたら」 scores 2. Two hands, one English marker, opposite codes. The mark crossed in both and by different channels, and the instrument sees it in one. §10's measured losses have always been losses on the grammatical channel; this is the first case in the section of a gain the channel cannot see.
• A practising translator's own account of where the source decided for him. The lead's translator's log, frozen before the study existed, classifies 4 of 23 scene-B sites as the English words fixed the level and 19 as nothing did, and I supplied it. Three of the four are the same thing — a bare imperative. Against the two readers: determinate at 2 of 4 on the lead's FIX set and 1 of 19 on its SUP set (Fisher exact P = 0.067, direction as predicted, not significant, and reported as descriptive because 4 sites cannot carry a confirmatory claim). The translator and the readers are pointing at roughly the same small handful of places, and it is a very small handful.

What §10 may now say, and it is narrower than what §10.6 hoped for.

S1-c. Where a source marks social footing in no grammatical slot and the target must mark it in every utterance, the translator is choosing at nearly every site, not at a special subset of them. On 70 Dickens utterances, two independent readers of the English could jointly identify the footing from the words alone at 9. The advice "watch the sites where the source is silent" is not actionable, because competent readers do not agree on which those are; what is actionable is the opposite and duller instruction — assume you are supplying the relation everywhere, and decide it deliberately once for each dyad rather than sentence by sentence.

Limits that bind it: two readers, not many; both are model seats and neither is a reader in the charter's sense (§4); one work, one language pair, one published human hand; the coder is one axis of footing and not footing; and the three predictions this run was built around are withheld, so nothing here rests on them. Tier D is NOT PASSED.

7.39 Two translators printed the same refusal of the rhyme; only one of their pages keeps it — and an expanding translator adds clauses, not words inside them — 2026-08-27 (S228), RS-20260827c-declared-page

§7.37 found that a translator's printed refusal of a device predicts what a reader loses while his printed promise does not. This is the page-side half of the same question, and it is not a reader measurement: it asks whether the refusal is visible in the prose itself.

Two hands printed the same refusal of al-Ḥarīrī's rhymed prose — Chappelow in a footnote in 1767 ("nor shall I imitate the author in my translation"), Preston in his introduction in 1850 ("rhyming prose is extremely ungraceful in English"). Chenery 1867 printed nothing. Each hand's chime rate at adjacent clause-ends was divided by its own bearer-permutation null — that text's unit-final words shuffled 2,000 times, so a hand whose English is full of -tion endings gets no credit for what its word-list produces by accident. A ratio of 1.00 is exactly as chiming as its own vocabulary gives by chance.

Chappelow is at or below chance in all four cells. Preston is above chance in three of four, and above the hand who declared nothing in both U-COMMA cells. The lead's own arms on the same Arabic set the scale: 5.994 chasing the chime, 3.211 neither seeking nor checking, 1.002 de-chimed on purpose, 0.825 refusing and checking every pair. Chappelow's page sits where a hand that checked sits; Preston's sits where a hand that declared and did not check sits.

What the handbook may now say. A printed refusal of a device is not a description of the page. Of the two English translators who printed the same refusal of Arabic rhymed prose, one page carries it and the other does not, and the difference is 3× in a measure the reader never sees. If you mean to refuse a device, the declaration is not the act — checking is, and the check is cheap: grade your own clause-ends against your own word-list.

A second result, on the other thing a translator can do with a difficult text. Chappelow's page does something he never declares: he unpacks the commentary into the running text, at 5.30 English words per Arabic prose token against Preston's 2.07 and Chenery's 2.14. The registered question was where the extra words go, and the answer is that they arrive as new clauses. His mean comma-clause is 7.56 and 8.86 words against Chenery's 7.70 and 7.49 — his clause is an ordinary English clause — while he writes 2.30 and 2.29 clauses per Arabic colon against Chenery's 1.04 and 1.03. Expansion, when a translator explains inside the text, is realised in clause count and not in clause length.

And Preston's ceiling survives a third hand. His 95th-percentile comma-clause is 11 and 12 words against 15/15 and 15/20 for the two hands who declared no length rule (§7.35's finding, now with a second non-declarer), while his sentence is the longest or second longest of the three.

What §7 may NOT take from this.

1. No causal claim about declaring. Three fixed hands differ in person, date, edition and printing-house practice as well as in what they printed; the pre-run critic struck the causal wording before the run and the result page contains no causal sentence. P3 as registered — both declarers below the non-declarer — is CONTRADICTED, and what the handbook takes is the split between the two declarers, not the prediction.
2. The balance still belongs to nobody, and the equivalence claim also failed. P2 predicted the three hands' adjacency ratios would span less than 0.15 and it fails in 2 of 4 cells. The hand who printed the balanced period is the smoothest in one cell of four; the hand who printed nothing is the smoothest in three; Chappelow is at chance in all four.
3. Not a reader measurement of anything. RS-20260825c remains the record that the project's preference instrument could not say whether any of this is worth having.
4. One text carried a large repair. Chappelow's 1767 long-s is OCR'd as f, and a declared f→s repair replaced 21–24% of his tokens before the chime was graded; it fired on 0–2 tokens in every other text. No page images were consulted.

7.40 A radif standing in a genitive to its bearer cannot be carried, whatever its part of speech — and, given the two renderings that are left, three seats take the rhyme — 2026-08-28 (S229), RS-20260828-forced-half

ARM-radif step 2, and the arm closes resolved at 2 of 2. §7.36 said whether the Persian shape can be carried into English is decided by the radif's part of speech, and gave a prediction table that came out 6 of 6. This section adds the case that table cannot see, and then reports what happened when the two renderings the failure leaves were put to readers.

7.40.1 — the part-of-speech rule is necessary and not sufficient, and what defeats it is the CONSTRUCTION the radif stands in. Three ghazals were chosen and enumerated before any line was written; all three radifs are things an English clause can end on — a noun (دوست), a possessive pronoun (اوست), a negated possessive predicate (تو نیست) — and all three fail. Persian ezāfe and the Persian possessive are head-initial: the bearer lands immediately before the radif (جمالِ دوست, نظیرِ تو). The English of-genitive puts a preposition and an article in that slot — the beauty of the friend, an equal **of yours — and a function word cannot be a rhyme-bearer; the English possessive genitive reverses the order and destroys the shape outright. So: ask not only what part of speech the radif is, but whether it governs its bearer. If it does, the shape is unreachable in English however clause-final the radif itself may be.

7.40.2 — and a third failure mode, which is neither. ۹۴'s radif of his leaves the possessed noun in the pre-radif slot, where it is rhymeable. Nothing structural blocks the shape. It fails because the eight bearers — evening-dark tress, pine-swaying stature, standing, ruby lip, message, snare-like tress, desire, slave — have no English rhyme family in common; the best family reached covered three of eight. Three poems, three reasons, and only one of them is §7.36's.

7.40.3 — the two policies that are left have opposite costs, and a translator can price them before starting. Keep the tail (REP) and every line ends on an unstressed syllable, because that is what a genitive radif is; keep the chime (CHI) and every line ends on a stressed rhyme. The chime is bought with sense and the repetition is bought with rhythm: the three CHI renderings cost three logged strains — a metonym, a named emotion, an added assertion — and the three REP renderings cost none, because with the tail fixed and no sound to satisfy every line was free to be literal. REP also keeps what CHI cannot: Sa'di commits ایطا, repeating his own bearer, in two of these three poems, and only the repetition policy can reproduce it — the chime policy silently corrects the poet.

7.40.4 — choose the rhyme family from the hardest position, not the commonest sound. In all three poems the family that worked was the one containing a usable word for the least tractable sense: ‑AY for مجال (go away), ‑ED for پیام (a word he has said), ‑AIR for نظیر (compare). A hand that fixes the family first and then hunts will fail at the position that has no member.

7.40.5 — and the measurement, which is an honest null with a substantial level underneath. Nine windows from the three pairs, three blind seats, both presentation orders, five information conditions — nothing said; told exactly what the Persian does at the line-ends; shown the Persian of those very lines; told a true fact about the Persian that is not about the line-ends; and shown a different Persian ghazal matched to the first for length, script and form. 270 forced choices, 0 void. All four registered contrasts are unestablished and none is close: the told movement is +0.111 of an available 2.000 (P = 0.563), the shown-this-source movement +0.074 (P = 0.750), and nothing survives Holm with or without the three windows an equivalence gate flagged. What is there instead is a level that did not move: across all five conditions the seats take the chime-keeping rendering 170 times to 100. And the instrument that failed at step 1 works here — order consistency 0.852 at baseline against 0.548 on minimal pairs, first position 0.5296 inside its bar, floor 12 of 12.

What a practitioner can take from §7.40. Before you start a radif ghazal, ask whether the radif governs its bearer; if it does, you are choosing between the two halves, not carrying both. Expect the repetition to cost you your line-endings and the chime to cost you your literalness. On this evidence — three model seats, not readers — the chime-keeping rendering is the one taken, and nothing you can put in a note or on a facing page changed that.

What §7 may NOT take from this.

1. Not that repetition is worth less than chime. The REP policy forces an unstressed line-ending at every position in two of the three poems, and the seats' own reasons name that trade in both directions. The level is a comparison of two policies as realised by one hand in English, not a measurement of a device. The confound was declared in the frozen design.
2. Not that disclosure does nothing. Four unestablished contrasts at nine clusters are four unestablished contrasts, and the design registered in advance that a non-significant result here is reported as unestablished, never as no effect.
3. §7.38 is not impeached, and its shown half is strengthened. §7.38's own limit conceded that its shown condition was longer and more salient than its comparators; the length-, script- and form-matched decoy used here removes that objection and the shown contrast is still null. Its told half did not reproduce — and the two disclosures are not the same object: S227's named a feature one arm plainly has and the other lacks; this one names two features and says outright that neither translation has both. A disclosure that favours neither arm has nothing to move a choice with is a conjecture with a mechanism, carried in NEXT.md as a candidate, and it is not part of this section.
4. Not that a reader prefers anything. Three model seats, Tier D NOT PASSED, every figure provisional and internal-judgment-only. Fourth section running to end at this sentence.

Evidenced on: Persian→English, one author, three poems, one hand, six whole renderings; nine constructed windows and three model seats for 7.40.5. 7.40.1–7.40.4 are craft findings from a frozen translator's log and carry no model judgment.

7.41 §7.36.1's and §7.40's impossibilities are REFUTED by two published hands — what survives is one prohibition, and it was not in this handbook — 2026-08-29 (S231), RS-20260829-radif-hands

This section corrects §7.36 and §7.40 and is the first test either has had against a hand other than this project's own. §7.36 closed with the scope line "Persian→English, one author, six poems, one hand", and that hand wrote the rule.

Two English translators of Hafez were coded on nine matched ghazals whose radif was enumerated mechanically from the Persian and classified blind by three independent seats (71 of 72 bits unanimous). Both books are public domain and were read whole.

7.41.1 — the transitive-verb cell is WITHDRAWN as an impossibility. §7.36.1's table calls a transitive radif "unreachable — the slot before the verb holds an auxiliary or an unrhymable subject". John Payne (1901) carries one at three of the three opportunities his volume reaches — می‌کرد as made, گیرد as taketh, دارد as hath — at every rhyming position of each poem.

7.41.2 — §7.40's genitive cell is WITHDRAWN as an impossibility. "A radif standing in a genitive to its bearer cannot be carried, whatever its part of speech." Walter Leaf (1898) carries one: turra-i mushk-sā-yi tu becomes jealous to match thy hair, for thee**, at eight rhyming positions of eight.

7.41.3 — what was mistaken for grammar was register. Both escapes are one move each and both are visible on the page. Payne's is inversion: the object, complement or adverbial goes before the verb so the radif can stand last — "Of us many a year made", "The market of lovely ones slack demand taketh". That order is available in an archaizing nineteenth-century verse English and not in a plain one, and ~~Leaf, writing plainer, never uses it and drops every one of those radifs~~ — WITHDRAWN, §7.43.1 (S233): true of the nine matched odes, false of Leaf's book. At his ode XVI he inverts at nine rhyming positions of nine and carries a negated finite verb radif, نمی‌رسد → attaineth not, by doing it. Leaf's escape is re-attachment: break the genitive and hang the pronoun on the clause as an adjunct, paying in the relation — thy becomes for thee. So the instruction a translator can act on is not "you cannot", it is "you can, at this price, in this register". ~~in this register~~ — THE REGISTER CLAUSE IS WITHDRAWN, §7.42 (S232, the same day): the inversion costs 2.407 points of ten in plain English and 1.926 in archaic, and an inverted archaic line still scores below a plain direct one. The instruction is you can, at this price, and the register does not discount it. §7.41.1 and §7.41.2 stand.

7.41.4 — one prohibition survives, and it is the one this handbook did not carry. A radif that is a case particle is refused by every hand that had the chance: Leaf drops را, Payne — who carries everything else — drops را, and the lead's own attempt on a second را ghazal enumerated four options before writing a line and found none. English has no object-marking word of any class to put in the slot; a preposition adds a relation the Persian has not got, a vocative cannot stand last, and a discourse tag installs a speaker. This clause was registered at S230 and deliberately kept out of v0.2; it goes in now. Its sibling clause does not: S230 also conjectured that English cannot end a clause on a bare copula, and Payne ends on is through a whole ghazal.

7.41.5 — and the cheapest case is one nobody had noticed. Where the radif is a light verb whose object is the qāfiya itself — غارت کرد is made a plundering, not plundered it — the two halves of the rhyming position are one constituent, English has the same construction, and the shape crosses whole with no inversion at all. The lead's T-hafez-ghazals-radif-R55-v1 carries کرد at 9 rhyming positions of 9 with the chime at 9 of 9 and no filler bought, on a ghazal whose nine rhyme-words are all Arabic verbal nouns. This is why the transitive/ intransitive axis was the wrong axis: what decides the case is whether the radif's object is the rhyme.

7.41.6 — how often the question arises, measured. 305 of Hafez's 495 ghazals (0.616) carry a radif, and the commonest are finite verbs and copulas — است, بود, دارد, کرد, را. On the rule as printed, an English translator of Hafez could not have carried the shape in most of the book. Payne did.

What this section does NOT say. It does not say Leaf could not have inverted — the design sees two hands and cannot hold register fixed, and the limitation is registered in RS-20260829-radif-hands §11.2. It does not compare declared against undeclared practice as a test: Bell 1897, who promises nothing about form, carries a radif in 0 of 32 poems, and the three hands translated different poems, so that zero is descriptive. And it says nothing about what a reader takes from a carried radif; Tier D is NOT PASSED and no reader was asked.

Evidenced on: Persian→English, one author (Hafez), nine matched ghazals, three published hands and one lead rendering, plus a 495-ghazal source census. 7.41.1–7.41.6 are coded mechanically off public-domain printed pages and carry no model judgment; the grammatical classification behind them was bought blind from three seats and is internal-judgment-only.

7.42 §7.41.3's register clause is WITHDRAWN — the inversion is costly in both registers and archaism does not buy it back — 2026-08-29 (S232), RS-20260829b-inversion-price

This section corrects §7.41, written into this handbook the same day, and it is the first time this project has held a register fixed while moving a word. §7.41.3 explained Payne's carriage of a transitive-verb radif by an inversion "available in an archaizing nineteenth-century verse English and not in a plain one", and RS-20260829-radif-hands §11.2 registered that no design separating capacity from register existed. One was built: eighteen bayts of Hafez rendered four ways — direct and inverted order crossed with plain and archaic lexis — with the archaic arms generated by script from the plain arms by a one-for-one word map, so nothing but the named tokens differs and no cell can be better written than another. Three blind seats, 386 calls, both registered primaries FAIL and they fail together.

7.42.1 — the inversion is expensive and the register barely changes the price. On a 0–10 within-idiom well-formedness rating the inverted arm loses 2.407 points in plain lexis and 1.926 in archaic. The registered interaction bar was 2.556 (1.500, escalated by the run's own noise rule); the measured interaction is +0.569, below the bar and below the instrument's own repeat noise of 1.278. On a within-lexis forced choice, 101 of 106 decisive answers reject the inverted order — 0.963 of them in plain lexis and 0.942 in archaic, a difference of two percentage points against a registered bar of fifteen.

7.42.2 — what archaism actually buys, and why it is not enough. Asked whether a line is well-formed within its own idiom, the archaic arms score higher than the plain: +1.056 at direct order, +1.537 at inverted. So the manipulation landed. But the diagonal is what a translator has to act on: the inverted archaic line (5.056) is still 0.870 BELOW the plain direct line (5.926). Hath does not pay for the inversion. A translator who inverts and archaises ends with worse-formed English than one who does neither — which is the practical content of the finding and the thing §7.41.3 denied.

7.42.3 — so the instruction changes again, and it is now shorter. Not "you can, at this price, in this register" but "you can, at this price, and the register does not discount it." Payne paid the price; §7.41.1's coding of his page is untouched and no model judgment bears on it. What falls is the explanation this handbook attached to his doing it.

7.42.4 — and there is no single price to quote. Across eighteen bayts of one construction the inversion costs between 0.000 and 6.000 points, median 2.000, and at three of eighteen it is free. "…for from the cup this thread its order has" costs nothing; "The man who in his hand a cup has / the sultanate of Jamshid for ever has" costs six. A handbook cannot quote one number for this move, and the one item-level distinction the design coded in advance — whether what is fronted is a direct object or an adverbial, OBJ median 2.000 against ADV 1.833 — does not explain the spread. The translator's log registered a different conjecture before the run (end-weight: cheap when the fronted phrase is short, dear when it is long and internally structured) and this design cannot test it.

What this section does NOT say. The archaic factor here is morphology — hath, thy, thou and the inflections they force — applied by script; "archaic register" and "the -th form of the final verb" are the same manipulation and cannot be separated. What fails is the morphology-carries-it version of §7.41.3, not Payne's whole register: his diction, metre and verse convention are untested here and a translator reproducing all of them might buy something this design cannot see. The order factor is still hand-written by the lead, so an inverted line he could not write well is indistinguishable from one English does not permit. And Tier D is NOT PASSED: no reader was asked anything, and the measure is not the naturalness sense — it is passage-indexed and therefore declared outside the six senses.

Evidenced on: Persian→English, one author (Hafez), two ghazals, one radif (دارد), 18 bayts in four machine-checked arms, three non-Anthropic seats, 386 calls, 0 dead; two pre-run critic seats, both NEEDS-REDESIGN, 30 findings, 9 BLOCKING. internal-judgment-only.

7.43 §7.41 and §7.42 gain the anchor they stood without — and §7.41.3's sentence about Leaf is CORRECTED on his own book — 2026-08-30 (S233), RS-20260830-leaf-contract

This section does two things and neither is a new recommendation. It gives §7.41 and §7.42 the Tier 1 evidence point D-20260724-04 requires — A-leaf-hafez, Walter Leaf's Versions from Hafiz (1898) read whole against the Persian, with his five printed clauses audited across all 28 of his odes — and it corrects one sentence of §7.41.3 that the fuller reading shows to be false.

7.43.1 — §7.41.3's clause about Leaf is WITHDRAWN. It reads: "Leaf, writing plainer, never uses it" — the line-end inversion — "and drops every one of those radifs." True of the nine matched odes S231 could see; false of the book. At his ode XVI (Brockhaus 266, radif نمی‌رسد na-mī-rasad, a negated finite verb) Leaf inverts at nine rhyming positions of nine and carries the radif by doing it: "Lo now, my heart to peace, as the years roll, attaineth not", for attaineth not to peace. This makes §7.41's refutation stronger. The inversion is not one translator's mannerism; it is used by the plainer of the two hands, in a third grammatical cell, in a book that prints a contract undertaking to reproduce the Persian form. Counter-instances to §7.36.1's impossibility now stand at five, in three cells — transitive finite verb (Payne, three), pronoun in ezāfe (Leaf), negated finite verb (Leaf).

7.43.2 — a printed self-assessment about frequency can be accurate, and §7.37 gains a third kind of statement. §7.37 established that a translator's printed refusal of a device predicts what a reader loses, and his printed promise to preserve one leaves no trace. Leaf prints a third thing: a comparative self-report — he repeats a rhyme-word "rather more sparingly than Hafiz". It holds. 0.050 against Hafez's 0.089 on the same 23 poems, below on 11, above on 3, level on 9 (exact sign test P = 0.057; P = 0.022 excluding one ode whose refrain the OCR corrupts). The project registered in advance that it would fail, on the ground that English has the poorer rhyme inventory, and the registered prediction is wrong. A handbook reading prefaces as evidence should treat what I do more or less often than my author as a checkable claim, not as modesty.

7.43.3 — what a strict metrical contract actually costs, exhibited rather than measured. T-hafez-bekonad-R57-v1 renders Hafez غزل ۱۸۷ — Leaf's own ode XV — twice under R57, Leaf's contract, with the metre the only clause that varies. The radif survives the metre: eight rhyming positions of eight close on will do, at fifteen syllables a line, fourteen lines running, which is an existence result at the one item of the forbidden cell where the published record has no counter-instance. What the metre takes is content words of the wrong shape — thirteen losses in fourteen lines, twelve of them lexical: the numeral صد twice in consecutive bayts where Hafez repeats it, the epithets پری‌چهره and مشفق, the realia فاتحهٔ صبح (which becomes a dawn-psalm, a change of religion), a root-play. Not one loss is caused by the rhyme or the radif. So the instruction is: a fixed measure is paid for in nouns and numerals, and a translator signing one should decide in advance which of his author's small words he is prepared to lose. Craft finding from a frozen translator's log; one hand, one poem, no model judgment.

7.43.4 — the line-end is a fixed quantity of room, and two hands spent it differently on the same line. Leaf's ode XV rhymes on -end; this rendering, written without seeing his, rhymes on -ending. Same poem, same metre, same English rhyme family, opposite purchase:

Leaf 1898: The bitter pray'r of the midnight a hundred ills shall amend. R57: One prayer at night, in the deep dark, all ills' defending will do.

He puts the doing into the rhyme-word and loses the device; the other arm puts the rhyme-word before the device and loses the numeral. This is an exhibit, not a measurement, and it is the clearest one this handbook holds of what a line-end costs.

What this section does NOT say. It does not rank the two hands and it makes no quality claim — Tier D is NOT PASSED. It does not say Leaf keeps L-METRE: what is measured is the necessary condition, that the ode's modal English syllable count equals the syllable count of the Persian measure his own table names (21 of 28 exactly, 28 of 28 within one syllable, six of the seven misses in the direction his own instruction to "dock the smaller parts o' speech" predicts). Whether he holds the long/short pattern is unmeasured. And every figure about Payne 1901 is withheld: the OCR of his volume runs footnote prose into the verse and no rule this session could write separates them, so RS-20260829-radif-hands §4 remains the only coded Payne evidence.

Evidenced on: Persian→English, one author (Hafez), one hand read whole — 28 odes, 444 printed lines, 250 rhyming positions; the source censused at 495 ghazals; two pre-run critic seats, both NEEDS-REDESIGN, 28 findings; blind rhyme adjudication with a mechanically-certain calibration set. internal-judgment-only.

7.44 The rhyme sense does not sit at the English rhyme — it sits one word to its left, and §7.43.2's third kind of self-report gains a fourth that is FALSE — 2026-08-30 (S234), RS-20260830b-rhyme-family

ARM-rhyme-family step 1 (T2). The registered primary is WITHHELD by two of its own gates, and what withheld it is the finding. The conjecture under test was written into a translator's log at S233 and had never been checked: that whether a ghazal can be monorhymed in English is decided by whether the source's rhyme-bearing senses happen to lie inside one English rhyme. Two quantities were built for it on Leaf's twenty-eight odes, from disjoint seats at disjoint labs — available coverage A from the Persian alone, realised coverage R on Leaf's English — after two independent critics named shared-rater variance as the design's first fault.

7.44.1 — a published English monorhymer's rhyme words almost never carry the senses his author put at the rhyme. Pooled realised coverage across Leaf's 28 odes is 0.0384 — nineteen odes at exactly zero, the best at one in six. The same blind answers, with the requirement that the named English word be the rhyme removed, give 0.311, eight times as much. So the senses are in the line-endings and they are not at the line-end: they sit in the word immediately before the rhyme, and the rhyme is spent on whatever will rhyme. This is measured on one hand and it is the first time this project has separated the sense reaching the end of the line from the sense reaching the rhyme.

7.44.2 — §7.43.2's kinds of printed self-report gain a fourth, and it is the one that fails. §7.37 had refusals (predictive) and promises (untraceable); §7.43.2 added the comparative self-report about frequency, which held on Leaf's own book. This adds the translator's account of what his chosen rhyme is doing — and on a blind reading it is wrong. T-hafez-bekonad-R57-v1 §D2 says its -ending family "happens to hold, in one rhyme, a word for almost every sense the poem asks for". Shown that rendering's eight line-endings and the Persian rhyme words, a seat that had no part in writing it found the senses at five of seven positions and named things, calamities, cruelties, God, prayer — every one of them the word before the rhyme, with amending, defending, ending, sending, ascending taking the line-end. The family did not supply words for the poem's rhyme senses; it supplied verbs that could govern them. The sentence in that log is corrected here rather than in the frozen log, which stands as written.

7.44.3 — availability is close to constant, so it cannot be what decides. A across 28 Leaf odes and 12 unselected ghazals lies between 0.09 and 0.38, median ~0.22 in both sets: English gathers about one rhyme sense in five into a single rhyme, in every ghazal tested. The commonest covering rimes are -ite and -ation. And Leaf did not choose his poems for it — mean A 0.2215 for his twenty-eight against 0.2231 for twelve ghazals nobody chose, Mann–Whitney U = 161.5, two-sided P = 0.855. The S233 log's picture of lucky and unlucky poems is not what the measure sees; غزل ۱۸۷ was a lucky hand, not a lucky poem.

7.44.4 — what a practitioner can take, and it is smaller than §7.36's. If you are carrying a monorhyme into English, do not expect the rhyme to hold the source's rhyme sense, and do not promise that it will. Expect to place that sense immediately before the rhyme and to spend the line-end itself on the rhyme. This is a description of what one published hand did at 250 positions and what one other hand did at seventeen; it is not yet advice with a second hand behind it, and §7.44.5 says why it may never become one.

7.44.5 — two things this section may NOT be read as saying. It does not say the conjecture is refuted: P1 returned ρ = 0.3136 at one-sided P = 0.05212 on n = 28, which is powered only for ρ ≈ 0.45, and the run's own placebo gate fired (mean qāfiya character length predicts R at ρ = 0.4064, better than A does), so no mechanism claim is made in either direction. It does not say A is a property of English: A is a lower bound found by one seat asked for eight words a sense, and the result page forbids the stronger reading.

Evidenced on: Persian→English, one author (Hafez), one published hand at 28 odes and 250 rhyming positions, plus two lead renderings under one arm of R57; one rater per side, disjoint labs; twelve unselected ghazals for the selection null. Two pre-run critic seats, both NEEDS-REDESIGN, 27 findings, eleven amendments accepted and three remedies overruled in writing. Verifier 102 checks, 0 failures. provisional; no sense Tier-D calibrated. Limits in RS-20260830b-rhyme-family §7, including a stated blinding breach on the translation limb.

7.45 When a source changes language and the target cannot, the choice is not whether to mark but whether to keep — 2026-08-31 (S235), RS-20260831-arabic-in-persian

ARM-gulistan span C. «گلستان» is bilingual: Sa'di writes Persian and quotes Arabic, and a Persian reader sees the language change on the page. 21 Arabic quotations were enumerated from the Persian across باب دوم before any English was opened; the primary count is the 15 in حکایات ۱–۳۱, coded in four published English hands, 58 of 60 cells aligned.

7.45.1 — half of these translators supply a signal the author did not. At the loci Sa'di does not frame, the pooled marking rate is 19 of 35 = 0.543 — Eastwick 1852 at 11 of 12, Ross 1823 at 8 of 10 — against a registered prediction of ≤ 0.35. At the loci he does frame it is 9 of 12 = 0.750. The marking is typographic, not verbal: no hand in the count wrote in Arabic or the Koran says where the Persian had not.

7.45.2 — the marking can be finer than the line. At the one locus where a Qur'anic hemistich rhymes with the Persian hemistich before it, Eastwick prints the Persian half roman and the Arabic half italic inside one couplet, and his ordinary verse in the same book is roman. He does not carry the rhyme; he carries the language change. The registered prediction that no hand would mark this locus is FALSE.

7.45.3 — the operative rule, and the only one here that is unanimous. Where an author quotes a foreign tongue and glosses it himself, the English says one thing twice unless something marks the first. The two hands that mark the Arabic keep the doubling; the two that do not delete the Arabic and print the gloss alone — four of four. So: decide whether you will mark before you decide whether to keep, because the second decision is made by the first. A translator who will not mark should delete rather than print the repetition, which is what both refusers did.

7.45.4 — what a uniform "no marking" can mean. Gladwin 1806 marks at 0 of 12 unframed loci — but his book uses italic only for running heads, so his zero is partly a fact about the volume's typography rather than about his policy. A hand's silence is only evidence of a choice where the hand had a device to use.

7.45.5 — the signal is about the language, not about the status of what is said — 2026-09-02 (S240), RS-20260902-arabic-function. The six loci deferred by the count above were coded, and all 21 were sorted by what the Arabic is doing: scripture, maxim, or verse quoted for what it depicts. Across the two hands that mark, the marking rate is 1.000 on the fourteen ornamental cells, 0.929 on the fourteen scriptural, 0.800 on the ten gnomic — and Ross italicises an Arabic couplet about pomegranate blossom while leaving a Qur'anic hemistich plain. In one tale, four lines apart, Eastwick sets Persian couplets roman and Arabic couplets italic with nothing else changing. So a translator who marks scripture and not ornament is telling the reader something the author did not. Registered prediction to the contrary: failed at both ends.

7.45.6 — a foreign quotation with nobody to attribute it to is the one thing this tradition will not print. Of the 21 loci, exactly one is deleted by all four hands — an unattributed, unframed Qur'anic rhetorical question dropped between two Persian couplets — and exactly one is left plain by all four, a definition that continues the argument instead of standing beside it. Both are places where the Arabic is not doing the work of a quotation. At the deleted locus the live options in published practice were mark it or cut it; printing it plain, which is what the source does, is a thing no hand did. Four hands, one book, one locus each way.

Evidenced on: Persian→English, one author (Sa'di), one chapter of one book, four published hands, fifteen loci, coded by the lead from uncorrected OCR with italic recovered from each scan's ABBYY layer. A second pair was added at S241 (§7.50, RS-20260903-latin-in-italian): Latin inside Italian, «Vita Nova» whole, four English hands, 21 loci. §7.45.1 and §7.45.6 travel; §7.45.3 is scoped to option sets in which the third move is a cut; §7.45.5's registered form fails on that material and survives only on a repaired measure — read §7.50 before applying any of §7.45 outside Persian. A third pair was added at S243 (§7.52, RS-20260904-french-in-russian): French inside Russian, «Война и мир» I.i.I–II, three English hands, 38 loci, 96 cells. It changes the scope of BOTH sections: §7.45 and §7.50 are evidenced on a source that QUOTES another language, and on a source that CODE-SWITCHES the option set is different in kind — retention falls from 0.83 to 0.156 and the supplied signal stops being typographic. Read §7.52 before applying either section to a source whose characters simply speak the other language. Two of 60 cells unaligned and excluded; the primary was narrowed from 21 loci to 15 after the design was frozen, by position in the chapter rather than by outcome, and the narrowing is declared. Unanchored coding, one reader: internal-judgment-only, every marked cell quoted verbatim in coding.md. provisional; no sense Tier-D calibrated. Limits in RS-20260831-arabic-in-persian §5.

7.46 The impossibility is withdrawn, not downgraded — a whole book of one hand against a whole book of the original — 2026-08-31 (S236), RS-20260831b-radif-hands

ARM-radif-hands step 2 (T5), and the arm closes here. RS-20260829-radif-hands refuted §7.36.1's printed impossibility on nine matched items and downgraded it to costly. This is the powered version: John Payne's volume 1 of 1901 is the whole of Brockhaus I–CC — 199 odes printed, XIV omitted as spurious by Payne himself — and he did not choose his poems. 173 of the 199 were matched to the Ganjoor census by two seats at disjoint labs, agreeing on 189 and recovering all ten held-out Brockhaus anchors, with the match independently checked by the rank correlation the two editions' shared alphabetical order forces (ρ = 0.876). Carriage was coded mechanically before any match existed, and a further two blind seats were asked whether each English tail actually renders the Persian radif.

7.46.1 — §7.36.1's transitive-verb prohibition is WITHDRAWN, and the cell it called unreachable is the one this hand carries most. Among 173 accepted odes, 137 have a census radif and Payne carries a repeated English tail at 110 of them, 0.803 (0.547 on the stricter measure where two blind seats confirm the correspondence). By radif class: finite transitive verb 32 of 38 = 0.842 [0.696–0.926], everything else 27 of 34 = 0.794. انداخت → did cast, کرد → made, می‌کنند → do practise, آورد → hath brought, نمی‌گیرد → takes not, دارد → hath. The rule is not narrowed; it is deleted, and what stands in its place is §7.41's and §7.42's register clause.

7.46.2 — §7.40's pronoun-in-ezāfe clause is withdrawn as a prohibition and reported as a count. 1 of 3, below the 8-item reporting bar. Three items decide nothing, which is itself the scope line that clause never had.

7.46.3 — the case particle is the one prohibition standing, and its evidence line is now real. All five را ghazals in Payne's volume are refused, at every coding bar, all five classified CASE_PARTICLE: YES unanimously by three blind seats. With Leaf's one refusal and the lead's own first-hand refusal with the options enumerated, that is seven refusals, two published hands, zero carriages. English has no word of any class that does what را does.

7.46.4 — the mirror image, new: English manufactures a radif where the Persian metrics record none, and it does so for one reason. 9 of the 36 radif-free ghazals carry an English tail (0.250, [0.138–0.411]) — and nine of nine are Persian repeating something the definition of a radif cannot count: a bound copula (بنیادست, خطاست), a bound person suffix (می‌سپارمت, گیسویت), a bound plural (رهند), a bound stem (برآید), or the qāfiya word repeated whole (نکرد, دارد). Not one is an English refrain with no Persian warrant. A language that marks person, tense or the copula on the word forces English to externalise the mark, and the externalised mark is a radif in the translation that the source's own metrics do not record.

7.46.5 — §7.41's craft clause of two days ago is CORRECTED, and by the project's own second commission of the error it had just diagnosed. §7.41 said: where the radif is a light verb whose object is the qāfiya, the shape crosses whole into English and needs no inversion. It was written on غزل ۱۳۱, where that holds at every position. T-hafez-sahar-bolbol-R55-v1 rendered غزل ۱۳۰ — same radif کرد, same rhyme sound, the relation holding at four positions of eleven — in a declared plain register, graded the radif UNREACHABLE, and its frozen log concluded that the obstacle was lexical: English splits کردن into do and make and neither collocates with all eleven nouns. Payne's ode CXVI is that poem and he carries کرد as made at 9 of 9, including two of the three positions the log named as blocked, by nominalising the predicate and fronting the complement — made good actions her practice, made me to rue. The obstacle was the declared register, not the lexicon. §7.36.1 mistook one translator's register for a fact about English; this project made the same mistake four days after diagnosing it. The operational test is therefore not is the qāfiya the radif verb's object but does one English verb serve every collocation the radif makes across the whole poem, in the register you have declared? — and that is still answerable from the source before a line is written.

Evidenced on: Persian→English, one author (Hafez), one published hand across a whole volume — 199 odes, 173 matched, 137 with a source radif — against a 495-ghazal mechanical census; two matching seats at disjoint labs, three blind classification seats, two blind correspondence seats. Two pre-run critic seats, both NEEDS-REDESIGN, 22 findings, 6 BLOCKING. One hand and one register; Brockhaus is not Qazvini–Ghani; the accepted set is enriched for tail-carrying poems (0.688 against 0.577 among the dropped); classification unanimity fell to 0.880 and 34 of 87 radifs were dropped on it. provisional; no sense Tier-D calibrated; internal-judgment-only. Limits in RS-20260831b-radif-hands §8.

7.47 The metrical explanation of what separated two English Hafizes is dead, and the word-order one was withheld by its own gate — 2026-09-01 (S237), RS-20260901-inversion-habit

This section closes ARM-inversion-price and it closes it short. §7.41's explanation of why Payne carries the Persian radif and Leaf does not has now lost both of its clauses — the register one at §7.42, the one about Leaf at §7.43.1 — and this unit went after the two accounts that could replace it. One is dead. The other was measured and the measurement is not reported, because the design's own calibration gate admitted one seat of three where two were required.

7.47.1 — the ROOM account is dead, and it needed no model to kill it. Every printed hemistich of both books, counted with CMUdict and a fallback validated against 30 hand-counted lines: Payne's median is 15 syllables over 3,406 hemistichs, Leaf's is 14 over 444. The registered bar was a difference of two; the difference is one, and 0.87 after correcting for the counter's measured bias, whose differential between the hands is 0.133. Payne did not have room that Leaf lacked. Nothing in this handbook ever said he did — this closes the account before it could be written in, which is what a registered prediction is for.

7.47.2 — what a plain modern register can actually buy at the line end, from a frozen log. T-hafez-daasht-R58-v1 renders Hafez غزل ۷۷ whole under the new regime R58, in a declared plain modern English — no thou, no -eth, no hath — and carries the finite transitive radif داشت as had at 9 rhyming positions of 9. R58 requires the log to grade how far each inversion travelled, and the tally, computed by a script that parses the frozen log, is LOCAL 6 · SPREAD 3 · REBUILT 0. At six of nine only the verb moved. At three a second constituent moved with it. At none of the nine was the predicate reshaped — the nominalise-and-front move §7.46.5 caught Payne using (made good actions her practice) was never needed. So: a translator writing plainly can buy this inversion one line at a time, and the price is not a register — it is the subject–object juncture with no verb between them, which the log names as the hardest reading in the poem. One hand, one poem, nine positions: an existence result about a register, not a rate.

A second observation from the same log, and it is the mirror of §7.46.5's: داشتن has one English collocate that serves all nine of the complements Hafez gives it, so the lexical wall that stopped sh130 did not arise. The operational test §7.46.5 states — does one English verb serve every collocation the radif makes, in the register you have declared — was applied in advance here and it passed. The chime did not: HELD at 3 of 9, REFUSED at 6, because the English rhyme family that reaches two of Persian's nine -ār senses reaches no third without a word the Persian has not got.

7.47.3 — and the handbook is told, in writing, what it does not know. The HABIT account — that the line-end inversion is a standing property of Payne's line rather than a device bought for the radif — is unmeasured. A powered design was built for it, survived 28 critic findings, ran to completion on 523 items and three seats, and was withheld: one seat cleared both registered gates and the design required two. No inversion rate for either hand may be cited from this session. §7.42's instruction stands unchanged and unexplained: you can, at this price, and the register does not discount it.

What this section does NOT say. It does not say the two hands invert at the same rate, or at different rates. It does not say the inversion is local in general — §7.47.2 is one poem in one register by one hand, and the hand is this project's. And it makes no quality claim about anything: Tier D is NOT PASSED.

Evidenced on: Persian→English, one author (Hafez); §7.47.1 on two published hands read whole — 3,850 printed hemistichs — with a syllable counter validated against 30 hand-counted lines; §7.47.2 on one frozen translator's log, nine rhyming positions, no model judgment. Two pre-run critic seats, both NEEDS-REDESIGN, 28 findings, 9 BLOCKING. provisional; no sense Tier-D calibrated; internal-judgment-only. Limits in RS-20260901-inversion-habit §7.

7.48 A translator who signed for a quantitative measure delivered its shape, not just its length — and the thing that defeats English is a stress shortage, not a want of discipline — 2026-09-01 (S238), RS-20260901b-leaf-pattern

This section completes §7.43, which gave §7.41 and §7.42 their anchor and left one clause of that anchor's contract unmeasured. A-leaf-hafez recorded Walter Leaf's 1898 promise that "Each line of the ode from first to last is in one measure" and could show only that his English lines had the length of the Persian measure. Whether they had its arrangement is the clause. Two pre-run critic seats returned NEEDS-REDESIGN with 29 findings; the primary was replaced and one registered prediction was withdrawn as arithmetically forced before any figure was computed.

7.48.1 — the arrangement is delivered, and the null that shows it holds his own inventory fixed. On 301 lines across all 28 odes and 18 distinct measures, lexical stresses stand at the scheme's short positions at a rate of 0.0472. Permute only the arrangement of that scheme's shorts — its length and its number of shorts held exactly — and the rate is 0.4982. D = 0.4509, one-sided cluster permutation P = 0.0001, the metre as the cluster. All 28 odes positive. Against a strict iambic template of the same length D = +0.029; against his own prose, cut into 870 windows and scored on the same schemes, his verse is +0.3206 better [+0.2919, +0.3495]. Shuffling each line's own stresses collapses D to +0.007.

7.48.2 — so "in the original forms" is a claim that can be met, and one hand met it. Until now every clause this handbook has tested from a translator's preface was either a refusal (§7.37), a promise that left no trace (§7.26), or a frequency claim (§7.43.2). This is the first formal promise about verse shape that a hand kept across a whole book, and it is measurable without a reader. The instruction that follows is not about Persian: where a source form is defined as an arrangement rather than a count, a target-language rendering can be audited against that arrangement mechanically, and the audit has teeth — chance sits at 0.50 and a hand who tried sits at 0.05.

7.48.3 — what actually stops a stress language carrying a quantitative measure, in the terms its own translator gave. Leaf printed a diagnosis: "many Persian metres so abound in long syllables that the English language will not supply stress enough to reproduce them." Across the 18 measures, the longs a line could not have filled however placed rise with the measure's long-fraction at ρ = +0.831; the share of the stress a line has that he puts on longs does not fall (ρ = +0.228) and sits between 0.940 and 1.000 in every measure. The failure is a budget, not a scatter. Practically: before rendering a quantitative measure, count its long positions and count the stresses a line of your language can carry — the shortfall is arithmetic and is knowable before a word is written.

7.48.4 — and his examples are looser than his mechanism, which is the part to distrust. Leaf names metres 4 and 6 as the hard cases. Against the other sixteen they differ by +0.093 at P = 0.068, and metre 16 carries four longs to every short — more than either he names — and he does not name it. A translator's printed reason held; the printed list attached to it did not separate. Take the mechanism from a preface; re-derive the cases.

Scope and what this does NOT license. Persian→English, one hand, one book, 28 odes, and the coder reads CMUdict lexical stress plus a declared closed-class list, not the stress a reader performs — Leaf himself writes that "Stress is largely rhetorical." No sentence here says he reproduced the Persian rhythm as anyone hears it, and the project has no human readers to say whether he did. The estimand is conditional on lines that already carry the measure's exact syllable count, excluding 86 resolvable lines. Tier D is NOT PASSED and nothing above is a quality claim. The translation limb T-hafez-derakht-R57-v1 — غزل ۱۱۵ whole in metre 6 — is a labelled subject, not evidence about Leaf.

7.49 A rhyme costs the last word of the line, and it costs it for every sense — but the pre-rhyme slot §7.44 named is a fact about English, not about rhyme — 2026-09-02 (S239), RS-20260902-rhyme-slot

This section closes ARM-rhyme-family and corrects half of §7.44's headline on §7.44's own materials, with a second hand added. §7.44 established, on Leaf 1898 alone, that an English monorhymer's rhyme words almost never carry the senses his author put at the rhyme — 0.0384 across 250 positions — and read the residue as the sense sits one word to the left of the rhyme. That reading came from one poem, the lead's own. It now has two published hands, 54 ghazals and 964 sense positions against it, and it splits in two: the first half survives and grows, the second half is withdrawn. Two pre-run critic seats returned NEEDS-REDESIGN with 19 findings, and their central objection — that the previous design's predictor was a Persian lexical category and not an English rhyme obligation — is why this section can say what it says.

7.49.1 — the rhyme takes the line's last word, and it takes it from senses the source did not put at its rhyme either. The predictor here is a property of the English line: a line is rhyming when its own pre-radif content word carries the rendering's modal rime, and open otherwise. Take only the senses that come from a ghazal's non-rhyming hemistich-ends — ordinary Persian words, no qāfiya stock, no fused copula — and ask where they land. In an open English line such a sense stands on the line's last word 26.95% of the time (n = 141). In a rhyming English line, 2.44% (n = 41). The difference is +0.2451, cluster bootstrap interval [0.1735, 0.3129] over whole renderings. A factor of eleven, with the Persian word class held constant. It holds in both hands separately — Leaf +0.2115, Payne +0.2649 — and in both strata, renderings that carry an English radif (+0.2571) and renderings that do not (+0.2413).

7.49.2 — §7.44's second clause is WITHDRAWN. The pre-rhyme slot is where English puts a displaced sense-word, rhyme or no rhyme. Among carriers that are not on the qāfiya slot, the share standing exactly one word to its left is 0.1780 in rhyming lines (n = 118) and 0.1727 in open lines (n = 139) — against a uniform-placement null of 0.1045 and 0.1014. The concentration is real, about 1.7× uniform, and it is identical in lines that must rhyme and lines that need not; the two distance distributions have the same shape. So "it sits one word to its left" describes the penultimate slot of an English verse line and not anything the rhyme does. §7.44.1's sentence "they sit in the word immediately before the rhyme" is corrected here; §7.44.4's practical instruction is amended in §7.49.4.

7.49.3 — §7.44's first clause is confirmed, and the number it lacked is the one that matters. §7.44 had 0.0384 and nothing to compare it with. The comparison is 0.2695. What a translator loses by rhyming is not that his rhyme words happen to be poor at carrying sense; it is that the line-end slot is spoken for, and every sense that would otherwise have taken it is displaced. The cost is structural and it is priced: the last word of the line, about nine times in ten.

7.49.4 — what a practitioner can take, and it replaces §7.44.4's second sentence. If you are carrying a fixed rhyme into English, count the line-end as already spent: a sense that would naturally close the line will close it about a quarter of the time when the line is free and almost never when it must rhyme. Do not plan on placing the displaced sense anywhere in particular — the word before the rhyme is no likelier to take it than the word before an unrhymed line-end. Plan instead on where the line will still be free, and put the load there. Evidenced on two hands and 964 positions; the placement half is a null, and nulls are what the previous version of this advice was missing.

7.49.5 — three things this section may NOT be read as saying. It is observational: nothing was randomly assigned, both critics asked for matched unrhymed renderings by independent hands under assigned instructions, and the project cannot reach independent translators. It does not say the rhyme pushed a given sense out of a given line — 57.1% of located carriers stand outside their own bayt's two English lines, so the design deliberately makes no claim about which line a sense came from. And it does not revise §7.44's 0.0384, which was a differently-framed measurement on a different instrument; the two agree in magnitude and neither replaces the other.

Evidenced on: Persian→English, one author (Hafez), two published hands — Leaf 1898 at 28 odes and Payne 1901 at 28 — across 54 distinct ghazals and 747 sense positions, 368 located; plus a lead pair under the new regime R59, whose two arms differ in the rhyme clause alone, reported as description with its breach declared. One rater per stage, disjoint labs; a cross-poem control at 0.175 location against the true items' 0.492; a positional positive control that was void on its first execution and passed at exactly its threshold on a declared rebuild. Two pre-run critic seats, both NEEDS-REDESIGN, 19 findings, the primary estimand replaced and the permutation null withdrawn before any figure existed. Verifier 35 checks, 0 failures, three mutation tests, three caught. provisional; no sense Tier-D calibrated. Limits in RS-20260902-rhyme-slot §10.

7.50 A translator who marks a change of language marks it by contrast, not by italic — and where the reader can read the other tongue, nobody deletes: they translate — 2026-09-03 (S241), RS-20260903-latin-in-italian

ARM-latin-in-italian step 1 (T5). §7.45 is evidenced on Arabic inside Persian, one chapter, four hands. This is the second pair: Latin inside Italian, Dante's «Vita Nova» whole, 21 loci enumerated from the Italian before any English was opened, coded across four published English hands 1861–1902 — 84 cells, none unaligned. Three of §7.45's four clauses travel; one of them travels only after its measure is repaired, and the repair is the section's main contribution.

7.50.1 — the option set depends on what the reader can read, and §7.45.3 is a special case. Not one of the 84 cells deletes the quotation — KEPT 70, ENGLISHED 14, DELETED 0. In Persian the hands that refused to mark cut the Arabic and printed the gloss alone, four of four. In Italian the one hand who refuses to keep — de Mey 1902 — turns the Latin into English at thirteen of 21 loci and cuts nothing. So decide whether you will mark before you decide whether to keep (§7.45.3) is what the coupling looks like when the third option is a cut; give the reader the embedded language and the coupling runs to englishing instead, which costs the reader almost nothing. §7.45.3 is hereby scoped, not withdrawn.

7.50.2 — the signal is a CONTRAST; italic is only its commonest direction. Among retained runs the italic share is 62/70 = 0.886, which is §7.45.1 replicating. But the eight non-italic cells are not unmarked. Three of these four hands set Dante's divisioni in italic as a body and Martin extends it to the whole of chapter XXV, so a hand who wants to distinguish the Latin there must set it roman — which is exactly what Rossetti and Norton do at the Jeremiah locus and Martin at all six of Dante's classical citations. Measured as type-against-surrounding type: 69 of 70 retained runs are distinguished, 0.9857; 61 by italic and 8 by roman. The one undistinguished cell is Martin's Jeremiah, italic inside an italic divisione — §7.45.4's rule (a hand's silence is only evidence where the hand had a device) appearing as a mechanism. A study that codes "italic / not italic" will mis-score every locus whose surroundings are already italic.

7.50.3 — §7.45.5's registered form FAILS here, and its direction survives only on the repaired measure. The prediction was |italic-among-retained(dramatic speech) − italic-among-retained (scripture)| ≤ 0.15 per marking hand. Rossetti 0.2500 and Norton 0.2000 fail; Martin holds at 0.0000. Both failures are the single Jeremiah cell of 7.50.2. On the contrast measure Rossetti and Norton come to 0.0000 and Martin fails at 0.2000. What is safe to print without the repair is narrower: no hand here marks scripture and leaves dramatic speech unmarked, or the reverse — every difference between the classes is one cell wide and every such cell is explained by the type around it.

7.50.4 — what IS about status is a different decision, and it is not close. The hands agree on whether to mark. They differ, by function, on whether to translate the quotation for the reader: an English rendering is supplied beside the Latin at 21 of 21 retained cells where a figure in the book speaks Latin, and at 8 of 27 where Dante cites a learned authority (Fisher one-tailed P = 1.9 × 10⁻⁷). A translator marks the change of language wherever it happens; he translates the words where they are somebody saying something, and leaves them standing where they are a citation. The classification was frozen before coding; this particular contrast was chosen after seeing the table and is descriptive.

7.50.5 — §7.45.6 is the clause that travels best, and it survives a change of option set. The englishing hand keeps the Latin at 6 of 6 named classical authorities and at 2 of 15 loci elsewhere (Fisher one-tailed P = 0.0005). The one authority she englishes is Nomina sunt consequentia rerum — the only quotation in the book with no author's name on it. In Persian the one locus all four hands deleted was likewise the quotation with nobody to attribute it to. Two books, two centuries, two option sets, the same locus-type singled out: an unattributable authority is the one a translator will not leave standing in the other language.

7.50.6 — an author's printed language policy does not bind his translators, and may not even be read by them. Dante states one inside the book: he breaks off Lamentations after six words because «lo intendimento mio non fu da principio di scrivere altro che per volgare». All four hands print the sentence and all four break it — Latin at 21 loci, unglossed at 27 of 70 retained cells; Rossetti prints more of Virgil than Dante does; Rossetti and de Mey give Ovid's title in Latin where Dante gives it in Italian; and de Mey prints Dante's "the words cited are all in Latin" four sentences after printing those words in English. Where §7.48 and A-leaf-hafez found a translator's own printed contract kept, this is the other case: the author's contract, carried into the target and contradicted on the same page.

Scope amended S243: everything in §7.50 is about quoted Latin. On a code-switching source the clause that survives is §7.50.1's nobody deletes; §7.50.4's glossing finding becomes unaskable because the hands english instead of keeping. See §7.52.

Evidenced on: Italian→English, one author (Dante), one book entire, four published hands (Rossetti 1861, Martin 1862, Norton 1867, de Mey 1902), 21 loci, 84 cells, coded by one reader from uncorrected OCR with italic recovered from each scan's ABBYY layer, every cell quoted verbatim in E-20260903-latin-in-italian/coding.md. Two loci codings are arguable and named in RS-20260903-latin-in-italian §9; no claim above turns on either. The four hands are four books, not four independent draws. Translation limb T-vita-nuova-R04-v1 (XXIV–XXVI whole, 1,715 Italian words) is reported as a labelled fifth hand only, and its contamination gate came back DEPENDENT? against three of the four — it may not be used as an independent English witness. provisional; no sense Tier-D calibrated. Limits in RS-20260903-latin-in-italian §9.

7.51 The word-order account of what separated two English Hafizes is withheld a second time — and now for a reason: line-end order is at the edge of what blind coding can hold steady — 2026-09-03 (S242), RS-20260903b-line-end-order

ARM-line-end-order step 1 (T3), and it closes the arm. §7.47 left the HABIT account of what separated Payne 1901 from Leaf 1898 — is the object–verb inversion at the line end a standing property of a hand's line, or bought where a device needs it — unmeasured, its S237 figures withheld because one seat of three cleared the gates. This session rebuilt the measurement to answer the charge that S237's gate had been loosened after it fired: every voting code bought fresh, both gates binding, the synthetic floor restored, one defective calibration key corrected in advance, and a third seat added that had never coded the construct. Two pre-run critic seats, NEEDS-REDESIGN, 28 findings, 8 BLOCKING on that one point, all granted.

7.51.1 — the same result, and it is now evidence about the construct. Again one seat of three cleared both gates (qwen/qwen3.7-max, floor 23/24, gold 0.850), and again the primaries are withheld by a criterion registered before the deciding codes existed. The frontier seat that voted at S237 (gpt-5.6-terra) failed a fresh 60-item gold at 0.700 — its earlier pass did not replicate on independently drawn material. Two withholdings under two independent gates, one deliberately stricter, say the construct is codeable at about 0.80 agreement but not reliably enough for two certified seats. §7.47.3's "the project does not know" stands, and now names why.

7.51.2 — a method rule that binds any future count of line-end word order. Enjambment and apparent line-end inversion are confounded at the line end by construction: in the lead's own frozen gold, all 21 items whose clause continues past the printed couplet are coded INVERTED (against 21 of 39 among complete-clause items). A line that spills its clause forward reads as displaced whether or not the hand displaced anything. Separate enjambment from inversion before counting either. Method note (bsx).

7.51.3 — single-sample seat qualification does not hold. A seat certified by one small hand-coded gold failed a fresh gold drawn from the same construct. Qualify a seat on material its reported codes did not help choose, or the qualification is an artefact of the sample.

Evidenced on: Persian→English, one author (Hafez); 447 blind-coded line ends across two published hands and one frozen translator's log (غزل ۱۵, the no-radif counterpart of §7.47's sh77, rendered whole under the new regime R59). Three non-Anthropic seats; two pre-run critics, both NEEDS-REDESIGN. provisional; no sense Tier-D calibrated; internal-judgment-only. Limits in RS-20260903b-line-end-order §8.

7.52 Quoting another language and speaking it are different problems: where the author code-switches, translators do not mark the change — they english it, and the signal they supply is verbal — 2026-09-04 (S243), RS-20260904-french-in-russian

ARM-french-in-russian step 1 (T4). §7.45 and §7.50 are built on a source that quotes another language: Arabic inside Sa'di, Latin inside Dante. Both are bounded, attributable, announced. Tolstoy's French is none of those things. Nobody is cited; a character simply speaks French, and at five of the 38 loci a French sentence has Russian words standing inside it. Three published English hands, 1886–1923, 96 coded cells, two pre-run critic seats (NEEDS-REDESIGN, 31 findings, 11 BLOCKING, design rebuilt before any volume was fetched), and a blind second coder.

7.52.1 — the option set is different in kind, not in proportion. Dante's hands kept 70 of 84 cells (0.83) and deleted none. Tolstoy's hands keep 15 of 96 (0.156) and english 81 (0.844), and again delete none. The readership does not explain it: Victorian English readers of Tolstoy are Victorian English readers of Dante. What differs is that a quotation is somebody else's words and a code-switch is the speaker's own. A translator will leave another man's sentence standing in its language; he will not leave his character's speech standing in it.

7.52.2 — §7.50.1's clause is the one that travels, and it now stands on three pairs. Zero deletions in 96 cells, after zero in 84 and a measured cut-or-mark option set in Persian. Where the reader can read the embedded language at all, nobody removes what was said.

7.52.3 — englishing IS glossing, moved off the foot of the page and into the text; and this is the practitioner's real choice. Tolstoy prints a dual text — French in the body, his own Russian gloss at the foot — at all 38 loci; the project has met no other source that arrives with the author's own translation attached. All three hands collapse it: 91 of 96 cells put the rendering in the running prose. And at the few places where they do leave French standing they print nothing beside it — 1 glossed retention in 96 cells (Maude's footnote to la femme la plus séduisante de Pétersbourg), against a registered ≥ 0.80 that fails on every hand. So the decision at a code-switch is not mark / keep / cut. It is carry the author's dual text, or collapse it — and collapsing it is not a failure to gloss; it is glossing by other means, at the price of the switch itself.

7.52.4 — the supplied signal changes register with the choice: typographic where the language is kept, verbal where it is collapsed. §7.45.1 found half a tradition supplying an italic the author never gave. Here the supplied signal is verbal — Maude writes "written in French", "said she in French", "still in French"; Bell writes "and in French". Four verbal signals against one gloss. A collapsed switch leaves nothing to set in italic, so a hand who wants the reader to know a language changed has to say so in words. Garnett says nothing anywhere: her English of these two chapters contains no indication that a second language is being spoken at all. Practitioner's form: if you english a code-switch and the switch matters, the sentence saying so is the only device left to you.

7.52.5 — the second-order switch is the first casualty, and one hand shows it need not be. The five loci where Tolstoy puts Russian inside the French are englished by every hand that has them (retention 1 of 12 available cells). The single place in 96 cells where any hand registers the inner switch is Maude's, and he does it inside English, with quotation marks — "no longer my 'faithful slave,' as you call yourself" — and does not repeat it eighteen loci later where the same character uses the same formula. The device costs nothing and was used once.

7.52.6 — what could not be measured, and it is an instrument fact rather than a finding. The typographic prediction is void: the Project Gutenberg witness for the Maude carries no italic markup at all, and the ABBYY layer for the Garnett yields no italic in 1,164 pages. §7.45.1 could not be tested on this pair. Nothing here says these hands do not mark; it says the project could not see whether they do.

Evidenced on: Russian→English, one author (Tolstoy), two chapters of one novel, three published hands (Bell 1886 — from a French intermediary, by her own title page; Garnett 1904; Maude 1922–23), 38 loci, 96 coded cells, every cell quoted verbatim in E-20260904-french-in-russian/coding.py. A fourth hand (Crowell 1898) was dropped by a gate registered before it was opened: its scan names no translator. One hand's witness is missing six pages, coded NA-SCAN and excluded from every rate. The switch-type prediction is UNRESOLVED — it holds raw on all three hands but flips for one under the blind second coder, and its length-matched control passes on only one of the two hands that could carry it, so nothing in this section rests on it. Three hands are three books, not three draws, and two of them measure as dependent on each other. Translation limb T-voina-i-mir-R04-v1 (both chapters whole, 2,780 Russian words) is a labelled fourth hand only; its contamination is DEPENDENT? against all three. provisional; no sense Tier-D calibrated. Limits in RS-20260904-french-in-russian §8.

9. Changelog

• v0.2.52 (2026-09-04, S243), ARM-french-in-russian step 1 — §7.45 and §7.50 revealed as rules about QUOTATION. §7.52 new; the Evidenced on lines of both earlier sections amended; the arm closes resolved at 1 of 2. «Война и мир» I.i.I–II against three English hands, 38 loci, 96 cells, 0 deletions. Retention falls from Dante's 0.83 to 0.156 on a source that code-switches rather than quotes, so the option set is different in kind; §7.50.1 (nobody deletes) is the clause that travels, now on three pairs; englishing is glossing moved into the text — 91 of 96 cells — and only 1 of 15 retentions carries an English rendering beside it, against a registered ≥ 0.80 that fails on every hand; and the supplied signal turns verbal where the switch is collapsed (4 verbal signals, 1 gloss, no measurable italic). The typographic prediction is void on instrument grounds declared in advance, and the switch-type prediction is unresolved because a $0.12 blind second coder flipped it. No recommendation withdrawn; two scope lines narrowed.
• v0.2.51 (2026-09-03, S242), ARM-line-end-order step 1 — the word-order account withheld a second time. §7.51 new; the arm closes resolved. The HABIT account of what separated Payne from Leaf stays unmeasured, now with a reason: a rebuilt three-seat run (fresh codes, both gates binding, floor restored) again admitted one seat where two were required, so line-end word order is codeable at ~0.80 agreement but not reliably enough for two certified seats. Two method findings recorded — enjambment and apparent inversion are confounded at the line end (note (bsx)), and single-sample seat qualification does not replicate. No recommendation added or withdrawn.

• v0.2.49 (2026-09-03, S241), ARM-latin-in-italian step 1 — §7.45 taken off Persian. §7.50 new; §7.45's Evidenced on line amended to name the second pair and what did and did not travel. «Vita Nova» whole against four English hands, 21 loci, 84 cells, 0 deletions. Three findings the Persian material could not reach: the option set is mark / print plain / english where the reader can read the embedded tongue, so §7.45.3 is scoped, not withdrawn; the signal is a contrast (69/70 = 0.9857 distinguished, 8 of them by setting the Latin roman inside italic), so the italic proxy mis-scores every locus in an italic environment; and the decision that really tracks status is glossing, not marking — 21/21 for speech against 8/27 for cited authority, Fisher P = 1.9 × 10⁻⁷. §7.45.6 replicates in a different option set: the one quotation with no author's name is the one the englishing hand englishes. Registered P2 (§7.45.5's form) FAILS on two of three marking hands, on one cell each, and the repaired measure is marked post-hoc. P1, P3, P4 hold. Cost $0; no API call.

• v0.2.48 (2026-09-01, S238), ARM-leaf-contract step 2 — the arm closes resolved at 2 of 2, inside budget, both criteria met. §7.48 new. A-leaf-hafez's one unmeasured clause is measured: Walter Leaf's English delivers the arrangement of the Persian measure, not only its length. Stresses at short positions 0.0472 against a within-inventory permutation null of 0.4982; D = 0.4509, cluster-permutation P = 0.0001 over 301 lines and 18 measures, all 28 odes positive; iambic control +0.029, his own prose +0.3206 worse. Leaf's printed diagnosis of when English must fail is true in mechanism and loose in its examples: ρ(long-fraction, DEFICIT) = +0.831 while ρ(long-fraction, SUPPLY) = +0.228, so the failure is a stress budget, not a scatter; the two metres he names separate from the rest at only P = 0.068 and an unnamed metre is more long-heavy than either. Two critic seats, 29 findings, nine BLOCKING; the primary was replaced and one registered prediction withdrawn as arithmetically forced before any figure existed. Gate: 20 of 20 Ganjoor rows confirm S233's hand-transcribed metre table; metre 17 fails Leaf's own arithmetic and is excluded. Registered check Q7 FAILED — the coder found zero violations in the lead's own lines where at least one was predicted. Verifier 92 checks, 0 failures, 5 mutations, 5 caught. $0.117903400, all of it the critics.

• v0.2.47 (2026-09-01, S237), ARM-inversion-price step 2 — the arm closes short. §7.47 new. The metrical explanation of what separated Payne from Leaf is dead: their median hemistichs are 15 and 14 syllables, a difference of one against a registered bar of two, on a counter validated against 30 hand-counted lines. The word-order explanation was measured on 523 items across three blind seats and WITHHELD: one seat cleared both registered calibration gates where two were required, and the excluded seat had the run's highest agreement with the real-item key. §7.42 stands unchanged and unexplained. New craft clause from a frozen log: in a declared plain modern register the object–verb inversion carrying a finite transitive radif was LOCAL at 6 of 9 positions and required a predicate rebuild at none. $3.465563850, 100 calls; verifier 26 checks, 0 failures, 4 of 4 mutations caught.

• v0.2.46 (2026-08-31, S236), ARM-radif-hands step 2 — the arm closes. §7.46 new; §7.36.1 and §7.40 withdrawn as prohibitions; §7.41's craft clause corrected. A whole volume of one hand against a whole census of the original: Payne carries the radif at 110 of 137, and at 32 of 38 in the very cell §7.36.1 called unreachable. The case particle is refused 5 of 5, which with Leaf and the lead is seven refusals and no carriage. Nine of nine "manufactured" English radifs are Persian bound morphology the definition cannot count. And the project's own translation of the day, frozen first, concluded the obstacle was lexical and was refuted by Payne on the same poem: it was the register. $1.947399275, 55 calls, 6 lost to truncation (note (bsf), fifth firing, $0.301 wasted).

• v0.2.45 (2026-08-31, S235), ARM-gulistan span C. §7.45 new. Where a source changes language and the target cannot, half the published tradition repairs the signal the author did not give: unframed Arabic is marked at 0.543 against a registered ≤ 0.35, Eastwick at 11 of 12 and Ross at 8 of 10, and both by italic rather than by saying so. Eastwick marks at the hemistich, falsifying the prediction that the rhyme-bearing locus could not be marked. The one unanimous finding is §7.45.3: marking and keeping are one decision — the two hands that mark the author's self-glossed Arabic keep his doubling and the two that do not delete the Arabic outright, four of four. §7.45.4 records why a uniform zero is weak evidence. A-gulistan-hands gains §6 and its first erratum. $0 — no API call.

• v0.2.44 (2026-08-30, S234), ARM-rhyme-family step 1. §7.44 new. The registered primary is WITHHELD by two of five gates, and the floor that withheld it is the finding: realised coverage of a Persian ghazal's rhyme senses by a published English monorhymer's rhyme words is 0.0384 across Leaf's 28 odes, nineteen of them at zero, against 0.311 for the same blind answers when the named word need not be the rhyme — the sense sits one word to the left of the rhyme. §7.43.2's kinds of printed self-report gain a fourth and it FAILS: T-hafez-bekonad-R57-v1's log claim that its rhyme family held a word for almost every rhyme sense is wrong on a blind reading of its own eight positions. Availability is near-constant (A 0.09–0.38, median ~0.22 in both the chosen and the unchosen sets) and Leaf did not select his poems for it (P = 0.855). Translation limb T-hafez-amad-R57-v1, غزل ۱۷۶ whole under R57's FREE arm, chosen by the study limb as the lowest-coverage of twelve candidates and scoring above every one of Leaf's odes at the rhyme — description, not evidence, with a stated blinding breach. Predictor and outcome bought from disjoint seats at disjoint labs after both critics named shared-rater variance first. RS-20260830b-rhyme-family.

• v0.2.43 (2026-08-30, S233), ARM-leaf-contract step 1. §7.43 new. §7.41 and §7.42 now cite a Tier 1 anchor, A-leaf-hafez, instead of a result page. §7.41.3's sentence "Leaf … never uses it and drops every one of those radifs" is WITHDRAWN: his ode XVI inverts at nine positions of nine and carries a negated finite verb radif, نمی‌رسد → attaineth not — a fifth counter-instance and a third cell. Leaf's five printed clauses audited on all 28 odes: L-CHOICE holds decisively (18 of 24 metres covered; frequency-proportional draws reach 18 in 0 of 10,000), L-REPEAT holds and the registered prediction that it would fail is wrong (0.050 against Hafez's 0.089), L-METRE's necessary condition holds (21 of 28 exactly), L-RADIF is kept at the rate its hedge implies (7 of 28 odes). Translation limb T-hafez-bekonad-R57-v1, Leaf's ode XV rendered twice under R57 — the radif survives the metre and thirteen content words do not. Every Payne figure withheld on an extraction failure, declared. RS-20260830-leaf-contract.

• v0.2.42 (2026-08-29, S232), ARM-inversion-price step 1. §7.42 new, and it CORRECTS §7.41 the same day §7.41 was written. §7.41.3 said the line-end inversion that carries a radif is available in an archaizing register and not a plain one; the design §7.41 itself said did not exist was built — 18 bayts in four arms, order crossed with lexis, the archaic arms generated by script so no cell could be better written than another — and both registered primaries fail. The inversion costs 2.407 points of ten in plain lexis and 1.926 in archaic (interaction +0.569 against a bar of 2.556); on forced choice 101 of 106 decisive answers reject the inverted order in both registers. Archaism does raise within-idiom well-formedness (+1.056 / +1.537) and still leaves the inverted archaic line 0.870 below the plain direct one. New clause: there is no single price — 0.000 to 6.000 across eighteen bayts of one construction, and the OBJ/ADV distinction does not explain it. §7.41.1 and §7.41.2 stand; what falls is the explanation. Two critic seats, NEEDS-REDESIGN, 30 findings, 9 BLOCKING — one of which replaced the DV outright and one of which replaced hand-written arms with script-generated ones. Verifier 16 checks, 0 failures, 3 of 3 mutations caught; $1.208750650.

• v0.2.41 (2026-08-29, S231), ARM-radif-hands step 1. §7.41 new, and it CORRECTS §7.36 and §7.40 — the first time either has been tested against a hand other than the lead's. Both cells were stated as impossibilities and both are refuted by published pages: Payne 1901 carries a transitive-verb radif at 3 of 3 opportunities and Leaf 1898 carries a genitive one. What was taken for a fact about English is a fact about register — Payne inverts, Leaf does not — so the instruction becomes you can, at this price. One prohibition survives and it was not in v0.2: a radif that is a case particle is refused by every hand and by the lead's own enumeration, and S230's clause goes in; S230's bare-copula clause is refuted by Payne and does not. New craft rule: a light-verb radif whose object is the qāfiya crosses whole with no inversion. Census: 305 of 495 Hafez ghazals (0.616) carry a radif. Two critic seats, NEEDS-REDESIGN, 25 findings, 20 accepted, 3 in part, 2 overruled — the accepted A2 struck a gate that would have made refutation conditional on agreeing with the rule's author. Verifier 53 checks 0 failures, 3 of 3 mutations caught; $0.197706300.

• v0.2.40 (2026-08-28, S229), ARM-radif step 2, and the arm closes resolved at 2 of 2. §7.40 new, and it EXTENDS §7.36. The part-of-speech rule is necessary and not sufficient: a radif that stands in a genitive to its own bearer cannot be carried, because Persian's ezāfe and possessive are head-initial and the English of-genitive puts a function word in the rhyme slot — ۹۷ and ۱۲۶ fail that way and ۹۴ fails a third way, by exhaustion of the rhyme family. The two policies the failure leaves have opposite costs — the chime is bought with sense, the repetition with rhythm — and only the repetition policy can reproduce Sa'di's own ایطا. The measured half is an honest null: five information conditions, 270 forced choices, all four registered contrasts unestablished and none close, on a level that did not move — the seats take the chime-keeping rendering 170 to 100. The instrument that failed at S224 works on whole renderings: order consistency 0.852 against 0.548. No published figure is shown false; §7.38's shown half is strengthened by a length- and form-matched decoy and its told half did not reproduce, for a reason §7.40 states as a conjecture and does not adopt.

• v0.2.39 (2026-08-27, S228), ARM-balanced-period step 2, and the arm closes resolved at 2 of 2. §7.39 new — the page-side half of §7.37. Of the two English translators of al-Ḥarīrī who printed a refusal of his rhymed prose, Chappelow 1767's page keeps it and Preston 1850's does not: against each text's own bearer-permutation null, Chappelow runs 0.667 / 0.216 / 0.676 / 0.959 and Preston 1.585 / 1.106 / 1.680 / 0.465, with the non-declarer Chenery between them. Registered P3 is contradicted and the contradiction is the finding. Second result: an expanding translator's extra words arrive as new clauses, not longer ones — Chappelow's mean comma-clause is within 2% and 18% of Chenery's while he writes 2.2× as many per Arabic colon (P4 supported 2 of 2). §7.35's length finding survives a third hand (P1 supported 2 of 2); §7.35's balance finding is unchanged and P2's equivalence claim fails in 2 of 4 cells. RS-20260827c-declared-page / E-20260827c-declared-page; $0.14535690, all of it the pre-run critic, which returned two BLOCKING findings — the same one from both seats — and twelve findings accepted in full. Verifier 446 checks, 0 failures, 3 mutation tests, 3 caught. New regime R53, and «المقامة الدمياطية» translated whole under it. A-hariri-hands §4a corrected: its claim that Chappelow's book lacks the second Assembly is false.

• v0.2.38 (2026-08-27, S227), ARM-worth-paying step 2, and the arm closes resolved at 2 of 2. §7.38 new, and it splits §7.34: telling a reader the Arabic is rhymed prose moves preference toward the rhyme-forward rendering (7 of 54 within-cell moves, 0 against, Holm 0.0156, the segment-clustered check agreeing at 6–0); showing them the Arabic itself does not (6 up, 5 down), and the direct told-versus-shown contrast separates the two at 1 up / 7 down. A de-chimed control arm was built under R49 for the run. §7.34's both-orders statistic, registered in advance here, reads 3–2 and should not be cited again; its finding survives, its statistic does not. RS-20260827b-shown-or-told / E-20260827b-worth-paying; $1.631771425 against a declared $2.20; verifier 461 checks 0 failures; pre-run critic two seats, both NEEDS-REDESIGN, which moved the inferential unit and bought the control arm before any data call. Free and not ledgered: al-Ḥarīrī's second Assembly rendered whole twice more.

• v0.2.37 (2026-08-27, S226), ARM-declared-function step 2, and the arm closed resolved at 2 of 2. §7.37 new — a translator's printed refusal of a device predicts what a reader loses (three refusals, three confirmations); his printed promise to keep one leaves no trace three blind readers find at the 22 loci the study could enumerate, and one clear trace at a locus it could not. RS-20260827-declared-play. (Entry added 2026-08-27 by S227, which found the changelog missing a version while editing it; the figures are the result page's.)

• v0.2.36 (2026-08-26, S224), ARM-radif step 1. §7.36 new, and it CORRECTS §7.32. RS-20260826b-radif / E-20260826b-radif; $0.840041925 against a declared ceiling of $3.20; 204 calls, 0 dead, 0 re-dispatches, 0 unparsed; verifier 218 checks, 0 failures, 5 mutation tests, 5 caught; pre-run critic two seats on the design AND the exact texts — NEEDS REDESIGN, fifteen findings, seven BLOCKING, all accepted, none overruled, and two of them changed the manipulation itself. Translation limb T-saadi-ghazals-radif-R52-v1: six Sa'di ghazals whole, the project's first Persian ghazals — 39 bayts, 45 rhyming positions, depth 7–8, under the new regime R52, against every previous deep run in this project at depth 2–4. §7.32's sentence "English has no slot before a repeated word to rhyme in" is false and is struck: the slot exists wherever the radif is something an English clause can end on, and ۴۹ was carried at 8 of 8. The prediction table came out 6 of 6 from the radif's part of speech alone. The measured half established nothing and says so: all three registered predictions unestablished, two withheld by the run's own order criterion, and the diagnostic behind it is that three blind seats agreed with themselves across a swap of presentation order at 54.8% of cells against a chance rate of 50%. New note (brs). Contamination suspected, measured clean against the whole 68,000-word King 1926 Odes (0 shared 7-grams, longest run 5 tokens) but not against these poems, which the artifact says rather than glossing.

• v0.2.35 (2026-08-26, S223), ARM-balanced-period step 1. §7.35 new — Preston 1850's printed rule measured against his own printed page: the length half holds (230 lines across two Assemblies, longest 14 and 13 words, CV 0.198 / 0.177) and is imposed rather than inherited, while the balance half fails its registered primary. RS-20260826-balanced-period. (Entry added 2026-08-26 by S224, which found the changelog missing a version while editing it; the figures are the result page's.)

• v0.2.34 (2026-08-25, S222), ARM-worth-paying step 1. §7.34 new — the first time this project has asked which of two translations is taken, rather than how each is rated, and the first instantiation anywhere of the positive half of Preston 1850's sentence (R50, T-maqamat-sanaa-R50-v1, al-Ḥarīrī's first Assembly rendered whole a third time at $0). RS-20260825c-worth-paying / E-20260825c-worth-paying; 485 calls, 0 dropped, $1.787856750 against a declared $2.60; verifier 157 checks, 0 failures, 3 of 3 mutations caught; pre-run critic two seats, both NEEDS-REDESIGN, 28 findings, 15 accepted before any data call. The registered primary is WITHHELD by its own position-bias gate and the withholding is the section: with no context the seats take whichever passage they read first at 0.819, against 0.739 on a text compared with itself. The surviving measure is post-hoc and says the ranking of three policies inverts on one disclosed fact about the source.

• v0.2.33 (2026-08-25, S221), ARM-declared-function step 1, re-planned. §7.33 new — the first time this project has tested a published translator's printed account of what his device does to a reader, rather than an account it manufactured itself. RS-20260825b-flippancy / E-20260825b-flippancy; $0.616025750 against a declared $2.60; 279 calls, 0 of 135 scored cells dropped; verifier 141 checks, 0 failures, 3 of 3 mutations caught. Translation limb: al-Ḥarīrī's first Assembly rendered whole a second time under a new regime R48 "the rhyme first" — the artifact Chappelow 1767 and Preston 1850 both print a refusal to make — 33 full English colon-end rhymes against R43-v1's 2, contamination clean against all three published hands, plus the mechanical R49 control that removes the chime and nothing else. All three declared reader-effects fire in the declared direction and hold in each seat separately. Preston's exemption is WITHHELD by its own gate, and the withholding is the more interesting result: the comic half of the most comic maqāma is read as light subject matter 2 of 7 times, so the condition his exemption names is not instantiated by his own author. Two pre-run critics, both NEEDS-REDESIGN, both independently finding the same fatal defect — v1's primary item measured arch self-consciousness and not Preston's flippancy — 15 further findings accepted, 3 in part, 3 overruled in writing. Second session running in which the critic call that bought no data was the best-spent money on the page. Also: these seats returned 0 of 15 identical repeat vectors on a graded aesthetic judgement, against S220's near-determinism on a reference-resolution task — determinism is a property of the task, not of the seats.

• v0.2.32 (2026-08-24, S219), ARM-run-depth step 2, and the arm closes resolved at 2 of 2. §7.32 new — the first statement §7 has about a device's EXTENT rather than its presence, and the first answer to the question §7.28 had to leave open. RS-20260824c-run-placement / E-20260824c-run-placement; $0.477576775 against a declared ceiling of $1.20; 224 calls, 0 dead, 0 unparsed; verifier 103 checks, 0 failures, 5 mutation tests, 5 caught; pre-run critic one round, fifteen findings, all fifteen accepted, NEEDS REDESIGN on v1 — which rebuilt the experiment around two control placements and killed a rhyme that a heteronym had faked. The power gate passes (Δ at adjacent line-ends +0.321 against +0.25) and both registered predictions hold (Δ at the Persian's distance +0.214, ≥ +0.15 and less than the adjacent figure). Two seats of three carry the whole of it, and the extent gap lives entirely in the control cells — both bounds are on the section. Translation limb T-gulistan-bab1e-R47-v1, «گلستان» باب اول ۲۶–۳۱ whole under the new R47, which writes the same four lines at three placements at one sitting; the span's four-deep قطعه held clean, four distinct bearers, the project's first; contamination measured clean against Eastwick 1852 (0 twelve-grams, longest run 8 tokens). New note (brk): a mechanical rhyme grader that accepts any dictionary pronunciation can certify a chime a reader will not get.

• v0.2.31 (2026-08-24, S218), ARM-persian-hands step 2, and the arm closes resolved at 2 of 2. §7.31 new — §7.27 confirmed on a second author and a second language, and bounded. RS-20260824b-hariri-hands / E-20260824c; $0.483099750 against a declared ceiling raised from $0.60 to $0.90 by the critique; verifier 335 checks, 0 failures; pre-run critic one round, 14 findings, 5 BLOCKING, all accepted — among them the removal of a control that would have disqualified exactly the hands confirming the headline. The shelf gains A-hariri-hands, its second anchor on a source figure that sits in prose, its first on the work that defines the figure, and its earliest published hand at 1767. P1a (zero full rhymes) FAILS by one cell and P1b holds on every reading; P4 — COLA-KEPT ≥ 0.70 — holds at 0.974 and is the positive finding; P3 holds for one hand of three. Two translators' declared statements of policy, 1767 and 1850, are quoted on the anchor. Translation limb T-maqamat-sanaa-R43-v1, the maqāma whole — 139 prose cola and 9 verse lines — under R43, used on a second language for the first time; contamination declared suspected and measured clean against all three hands (longest runs 6, 7, 7 tokens), while Chenery and Preston measure closer to each other than either does to the lead. New note (brg).

• v0.2.30 (2026-08-24, S217), ARM-echo-threshold step 2 — entry written 2026-08-24 by S218, which found §7.30 in the body with no changelog row. §7.30 new. RS-20260824-eye-or-ear / E-20260824; $2.621514975; the half-chime is registered through the eye and only by a reader not attending to sound, the pararhyme alone is 0 of 28, and a NEEDS-REDESIGN verdict stands declared on the face of the section. Notes (bre), (brf).

• v0.2.29 (2026-08-23, S216), ARM-member-move step 1. §7.29 new, and it is a WITHDRAWAL of the handbook's last surviving positive instruction about sound. RS-20260823c-member-move / E-20260823c; 316 calls, 0 dead, $0.871405250 against a declared $3.40 ceiling; verifier 39 checks, 0 failures, 3 of 3 mutations caught; pre-run critic two rounds, 8 findings, 4 BLOCKING, six accepted and two refused in writing. F1 fires at 0.000 against a registered +0.10: widening the candidate pool from the source's rhyme-bearer to every content word of a plain rendering of the colon adds nothing at 50 STEM seat-loci, and nothing at the control either. §7.24 item 1's availability rationale is withdrawn, and its "not lexical, it is structural" clause is refuted by the translation limb: the hand's 12 chimes at 29 loci all use words absent from both blind renderings, and REORDER was used zero times in 42 attempts. What replaces it is a price: 12 carried, 10 refused, 5 nothing. Translation limb T-gulistan-bab1c-R45-v1 under the new R45, the span whole — 27 prose blocks and 42 bayts; contamination measured clean against Eastwick 1852 (longest run 7 tokens, 1 shared 7-gram), the first free published Gulistan comparator this project has reached. New note (brd).

• v0.2.28 (2026-08-23, S214), ARM-run-depth step 1. §7.28 new — the first section about a figure's EXTENT rather than its presence. «گلستان» باب اول ۶–۱۳ whole under the new R44, which forces the English rhyme to the source's own placement and depth. 9 of 9 قطعه runs held, 13 of 14 links full rhymes; depth four reached with four distinct bearers and depth five held only by repeating two — the English rhyme family runs out between four and five. Two seats reading the Persian flag the forced arm at 5 of 9 passages and the tradition's couplet arm at 4 of 9, on an instrument that caught 6 of 6 plants: the price of depth is repetition, not invention. The published census is 0 of 12 on both measures, six of the twelve cells being prose settings, and Eastwick writes an ababab sextet in the very passage where he answers none of Sa'di's rhyme-words. The reading primary SATURATED and is withheld by its own criterion 6 — 1.000 in both conditions — so nothing is claimed about reception. $0.842453050; verifier 260 checks, 0 failures; pre-run critic three rounds, 10 findings, 3 BLOCKING, all accepted. Note (brb).

• v0.2.27 (2026-08-22, S213), ARM-persian-hands step 1. §7.27 new. RS-20260822c-persian-hands / E-20260822c; $0.362263500 against a declared $1.30 ceiling raised twice during design; verifier 35 checks, 0 failures; pre-run critic three rounds, 13 findings, 4 BLOCKING, every one accepted and one remedy declined in part in writing. The shelf gains A-gulistan-hands, its first Persian anchor and its first record of what published hands do with a source figure that sits in prose. P1 holds and its stronger unregistered form is the headline: 0 of 128 cells graded STRICT. P2 fails on both clauses — the move into verse is taken at 4 of 128 — which refutes the natural inference from §7.26. P3 holds at +0.811 and +0.921. P4 holds at 20 of 32 after stage L, whose control caught 8 of 8 planted three-word additions. Translation limb T-gulistan-bab1-R43-v1, the Gulistan's first span past its preface, under the new R43; contamination DEPENDENT? against Eastwick at 13 tokens, outside the declared priming. New note (bra).

• v0.2.26 (2026-08-22, S212), ARM-echo-threshold step 1. §7.26 new, and it corrects §7.25 item 2's central sentence — published the same day on the lead's ear alone. RS-20260822b / E-20260822b; 773 bodies, 0 dead in the primary stage, $1.195830900 against a declared $1.90; verifier 332 checks, 0 failures, 4 mutation tests; pre-run critic two rounds, 10 findings, 4 BLOCKING, one refused in writing. A slant at a phrase-end is registered at +0.350 above a matched floor [0.183, 0.517], against a full rhyme's +0.578 — so "nobody hears these as echoes" falls as a general claim. It stands about its own three named pairs, which buy 0.000. The operative new distinction is position: +0.542 at a verse line-end against +0.222 at a prose colon-end, a move this handbook has been treating as one and which is two. Evidenced on Persian→English, 20 constructed loci from one work, three model seats, no human reader. The strongest caution is in the instruction itself: on a dedicated probe the seats called spelled-alike non-rhymes an echo at 0.722 against 0.778 for real rhymes spelled unlike, so §7.26 says registered, never heard.

• v0.2.25 (2026-08-22, S211), ARM-synonym-reach step 1. §7.25 new, and it is a WITHDRAWAL. §7.24 item 2 — enumerate the ordinary synonyms at a rhymed locus and look for a pair that chimes — is struck on a criterion registered before dispatch, tested on the project's first Persian work, Sa'di's «گلستان» دیباچه: two blind hands enumerating up to six ordinary renderings for each of 76 masked rhyme-bearers produce two full rhymes in 2,173 cross-member pairs, a yield of +0.042 against a bar of +0.25 and a withdrawal criterion of +0.10. What replaces it: the near-chime warning — relax the rule and reach goes to 0.708, and to 0.708 at the parallel cola the author left plain — and §7.24 item 1 promoted to the operative half, since the move that reaches an echo is structural and not lexical. 472 bodies, $1.067599850; pre-run critic two rounds, 8 findings, 4 BLOCKING; verifier 59 checks, 0 failures, 3 of 3 mutations caught. Chime decided mechanically by tools/rhyme_pairs.py, not by a seat.
• v0.2.24 (2026-08-21, S210), ARM-alf-layla step 7. §7.24 new. A registered test of the guess §7 had been carrying unregistered — that an Arabic rhyme reaches English only where the ordinary English words already chime — fails: 0.300 against 0.400, exact P = 0.452, and three published-hand counterexamples in Burton, one of which extends a three-colon rhyme to four by adding a colon. What replaces it is an instruction to enumerate synonyms rather than take the first ordinary word. 496 calls, $0.84032408 with the pre-run critic, verifier 199 checks, 0 failures; the pre-run critic returned 4 BLOCKING / 4 MAJOR / 1 MINOR and every finding was taken. Nothing is withdrawn from §7.18 or §7.22 except the assumption that a locus's members are fixed.

• v0.2.23 (2026-08-21, S209), ARM-matched-shape-heard step 1, and the arm closes resolved at 1 of 2. RS-20260821b-matched-heard / E-20260821b-matched-heard; $1.00069965 against a declared ceiling of $1.20, the runner's stop-loss firing three of ninety main cells short because the pre-flight omitted the reasoning cap (note bqs); verifier 185 checks, 0 failures, 5 mutation tests, 5 caught. §7.23 is new and it does not withdraw §7.22 — it withdraws the CHECK §7.22 invites. Translation limb: 劉基《賣柑者言》 whole from the Chinese, 312 characters → 449 English words under a new regime R40, which is R39 with the source-side class redefined for a language that matches by measure and R39's English resource table imported verbatim so the two censuses pool — 12 MATCHED, 0 IMPOSSIBLE, every locus exactly matched in English syllable measure, contamination 0/0/0, run 5. All twelve used resource (C), the matched frame, against the Arabic chapter's A6/B6/C2/D10 — noticed while rendering, not predicted, and reported as an observation rather than a prediction that held. Study limb: the subtractive test the two regimes' rule 5 seemed to have pre-built. Pre-run critic NEEDS-REDESIGN, 12 findings, 8 BLOCKING; v1 never dispatched; eleven accepted, one overruled on a stated ground (each dispatch is an independent completion with no conversation history, so a seat cannot recall its answer to the other version — its statistical consequence was accepted instead). Its BLOCKING 2 moved eligibility from the lead to an independent seat, and that seat killed the design: of 22 loci with intact material, the refused plain wording still echoes in form at 16, and a second seat — catching 4 of 4 planted content errors — says 10 of 22 do not say the same thing. Two survive both screens against a floor of ten declared in the design, and the primary was WITHHELD before a single main body was dispatched. What the run then showed is that the screens describe the passages: built matches recovered at 0.879 of cells (PR6, registered post-screen and pre-run, holds), and at the sixteen still-echoing loci 0.938 matched against 0.930 plain, a gap of 0.007. Even at the six loci the screener cleared, plain-arm recovery is 0.600 — the place is found either way, and what changes is what the reader calls it. New method notes (bqr) and (bqs).

• v0.2.22 (2026-08-21, S208), ARM-published-figure step 2, and the arm closes resolved at 2 of 2. RS-20260821-matched-shape / E-20260821-matched-shape; $2.097425 against a declared ceiling of $2.50 on a $5.00 day; verifier 248 checks, 0 failures, 5 mutation tests, 5 caught. §7.22 is new and §7.18 is corroborated rather than amended. Study limb: three blind seats, each given the whole 1819 chapter and a relation menu fixed before dispatch, re-code 14 of §7.18's 18 matched-shape loci and 40 loci of a third chapter — strict formal match at 0 of 14, 0 of 24 and 0 of 16, and the two loci where the lead's own coding had credited the published hand with a matched shape are overturned unanimously, in the direction that makes the published figure stronger. Translation limb: «باب ابن الملك والطائر فنزة» whole, the sixth Kalīla wa-Dimna chapter, 1,284 Arabic → 2,348 English under the new regime R39, which forbids a silent reduction at the matched-shape class — 24 MATCHED, 0 IMPOSSIBLE, at an expansion ratio of 1.829 against 1.849 for the same hand without the clause. Pre-run critic NEEDS REDESIGN, 3 BLOCKING / 7 MAJOR / 1 MINOR, all eleven accepted, none overruled: the English window abolished, the seat's threshold replaced by a reported relation, the falsify-a-published-figure-by-vote bar struck. New method notes (bqq) and (bqp).

• v0.2.21 (2026-08-20, S207), ARM-mimetic-subtraction step 1, and the arm closes resolved at 1 of 2. RS-20260820c-mimetic-subtraction / E-20260820c-mimetic-subtraction; $0.427041 against a declared arm ceiling of $1.20 and a session ceiling of $1.50; verifier 3,893 checks, 0 failures, 4 mutation tests. §7.19's instruction is WITHDRAWN and §7.21 is new. Study limb: on the same nine sites S204 subtracted the mimetic from by deletion, this run subtracted by grammatical substitution — mimetic → plain Japanese phrasing of the same event, written independently by two hands and agreeing at 8 of 9 under a rubric frozen before dispatch (note bqm). Primary denominator dropped from 9 to 7 by the pre-run critic (M08 and M12 excluded because each stranded a residual mimetic-derived form on the substituted sentence, recreating exactly the mutilation the run existed to avoid); the primary criterion was rewritten as a directional pair (≥ 5 forward AND ≤ 1 back) after the critic showed the S204 bar was not a valid sign-test criterion. Primary FAILS on both clauses: 1 forward, 1 back, 5 unchanged; p = 1.000 descriptive. And the mimetic-present baseline did not itself reproduce: 5 of 7 aggregate, 4 of 7 per item, for reasons unrelated to the mimetic (the seats' flag moved because of material in the English frame or a shifted attention to an omitted clause). Two readings of the null — the deletion flip was the mutilation and the plain substitute still carries the licensing property — are not separable by this design, but both withdraw §7.19's instruction. What survives: at the one site (N06) where the plain phrasing is genuinely more general than the mimetic, the English does read as an addition. The narrower test §7.21 states — write the plainest phrasing that carries no phonaesthetic form and ask whether your English still says more than that phrasing — fires at 1 of 7, not 7 of 9. One pre-run critic pass, NEEDS REDESIGN, 5 BLOCKING, 4 MAJOR, all answered; the stopping rule fixed at S206 held and no second round was bought. Translation limb: no new translation — chapters 2 and 3 of Botchan (frozen at T-botchan-ch2-R06-v1, T-botchan-ch3-R06-v1) are inherited whole; the arm is study-heavy by construction, with the substitutes standing in for the translation limb (this session's translation is the plain-Japanese subs, unbudgeted).

• v0.2.20 (2026-08-20, S206), ARM-device-function step 2, and the arm closes resolved at 2 of 2. RS-20260820b-device-function-2 / E-20260820b-device-function-2; $0.469376 against a declared $1.50 ceiling, 46% of it on three adversarial critiques; verifier 402 checks, 0 failures. §7.20 is new and it WITHDRAWS §7.16's subtractive-test instruction. Translation limb: the independent hand's fourteen replacement stretches, written by P1 as a budgeted contrast subject from the Chinese and the frozen English, blind to the labels — the lead's T-zhongli-R04-v1 held byte-identical, because holding it still is what the step measures against. Asked for the plainest rendering of the same events, that hand wrote a different device in: in great mouthfuls → greedily, at an unhurried walk → slowly, and the rhetorical question → another rhetorical question, so Δmanner is 0.000 at all three manner loci and every probe is 0.000 at the rhetorical question. What replaces the instruction is the narrower one that survives: write your replacement down and show it to someone else before concluding anything from it. And the finding under it — at two loci the plain version is read as MORE judging than the figure it replaced, because an adverb evaluates where a depiction shows. §7.16's attitude exception is restated as locus-specific by a branch fixed before the numbers were seen; its "a translator can tell when a choice is inert" survives (0.1111 against 0.4603). Three independent adversarial pre-run passes, all three NEEDS REDESIGN, 23 findings, 8 BLOCKING; v1, v2 and v3 were never dispatched. The second pass broke the design's own noise band and the registered stance prediction was rebuilt from scratch; the third removed every inferential word from the primary. A stopping rule against critic regress was fixed in writing before the third pass was bought, and applied as written.
• v0.2.19 (2026-08-17, S204), ARM-mimetic-carriage step 2, and the arm closes resolved at 2 of 2 by its first completion condition — a measured statement about carrying a target-less device. RS-20260817e-mimetic-reading / E-20260817e-mimetic-reading; $0.706588 against a declared $0.95; verifier 69 checks, 0 failures, 5 mutation tests. Translation limb: 「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 the two candidate renderings at every site written before the design existed. Study limb: the same English put twice to three seats, once against the Japanese and once against the same sentence with the mimetic deleted — 7 of 9 pairs flip from "adds nothing" to "ADDS", none flips back, p = 0.016 (p = 0.125 under the stricter of the design's two contradictory drop-rules, both reported). §7.19 is new: the English phonaestheme at a mimetic site is licensed by the mimetic and not by the event, and the practitioner's test is the subtraction — the reader-side mirror of §7.16. Two independent adversarial pre-run passes, both NEEDS-REDESIGN, 34 findings, 24 BLOCKING, all answered; v1 and v2 were never dispatched, and the second critic's objection — that a between-item null is not a null — invented the control that runs. The lead's own registered prediction 2 was refuted. No recommendation is withdrawn.

• v0.2.16 (2026-08-16, S201), ARM-device-function step 1, RS-20260816g-device-function. §7.16 is new and it adds the first §7 instruction that is not a count. Every instruction §7 has carried since §7.7 tells a practitioner to count marks, and §7.15 and §10.7 have each shown that the count's denominator is not a determinate set. §7.16 adds the subtractive test — write the passage without the device and ask whether the passage does the job anyway — which needs no denominator and can be run by one person on a draft. Evidenced on 蒲松齡〈種梨〉 translated whole, with the translator's account of each device registered in the frozen log before the design existed: the account holds on manner (3 of 3) and ornament (3 of 3), correctly calls three choices inert (0.0741 against 0.8642, exact P = 0.00455), and fails on attitude 3 of 3, where the attitude survives the subtraction at 1.00 and what the device was really contributing was the reader's belief that the source had a set phrase. Stated exception carried with the instruction: the subtractive test does not work on attitude, for the reason note (bpu) gives about depiction — the device and the proposition are the same thing. Read §7.16 with §7.14.

• v0.2.15 (2026-08-16, S198), ARM-supplied-footing step 2, RS-20260816d-lexical-channel. §10.7 is new and it closes ARM-supplied-footing resolved at 2 of 2, writing the subsection §10.6 has owed since 2026-08-15 — and writing it as a refusal plus one narrower instruction, because "the sites where the source is silent" turns out not to name a determinate set. §10.6's open item for that subsection is struck. Read §10.7 with §7.15: both are the same structural failure on different channels, a day apart.

• v0.2.14 (2026-08-16, S197), ARM-invented-ornament step 1, RS-20260816c-checked-ornament. §7.15 is new and it puts a condition under §7.14's diagnostic, written earlier the same day. Three seats shown the Arabic alone, one stretch marked, separate the frozen figure inventory's figured loci from its plain ones by 0.1886 against a registered gate of 0.40 — the gate fails and the run's source-visible arms are WITHHELD by their own rule. The failure is not incompetence: the seats agree with the inventory at its clearest cases (saj' chain 3 of 3, root-echo 3 of 3) and part company on three rules — unchanged-word repetition is not sound work to them ("repeated لسان is lexical repetition, not sajʿ, jinās, or muwāzana", 0 of 3), rhyme carried by a pronoun suffix is (3 of 3 at a locus the inventory excludes by rule), and matched imperfect verbs are not (0 of 3 at both). The two sets agree on 13 of 21 loci and on 5 of the inventory's 10 figures. So §7.14's diagnostic survives and its denominator must be declared: the devices the source licensed is a count under a rule, and the rule moves the answer. The positive instruction — compensate where the source has a figure — is refused a second time, now for a reason that owes nothing to any judgement of quality: its referent is not fixed. Pre-run critic NEEDS REDESIGN, 28 findings, 8 BLOCKING; twelve accepted and implemented, two overruled and registered as sensitivities, and one of the overruled ones was then reproduced by three seats that never saw it. Translation limb T-kalila-ibn-urs-R36-v1: a whole chapter under a regime that answers every enumerated figure and invents nowhere — 19 of 19 answered, 18 in kind, zero devices at non-figures, +35 words (3.3%) against the ornamentalist's +34 words (6.1%) for ten answers and twelve inventions. 351 bodies, $1.209261; verifier 48 checks, 0 failures, 5 of 5 mutations caught. Tier D still NOT PASSED; everything provisional.

• v0.2.13 (2026-08-16, S196), ARM-answering-figure step 2, RS-20260816b-invented-figure. §7.14 is new and it WITHDRAWS §7.12's item 3, published the day before. Two arms of one English translation, token-identical but for one word at eleven of thirteen loci, were read by three blind seats: the chiming arm is taken as evidence that the Arabic original had a sound figure at 9 of 13 places it has none, its de-sounded twin at 4 of 13, the plain base at 0 of 13 — so the supplied sound, not the matched members, moves the reader, and the seats say so in as many words ("only ordinary parallel phrasing"). And invented ornament at 9 of 13 is indistinguishable from genuine compensation at 8 of 13, so the inference carries no information about where the source's figures are. §7.12's item 2 survives and is explained: S191's invention arm was inaudible because it had no room, not because invention is inaudible. 402 bodies, $1.091013, verifier 596 checks, 0 failures, 5 of 5 mutations caught; all four third-party controls unanimous for a third session. Two pre-run critic passes, NEEDS REDESIGN then RUN WITH AMENDMENTS, 36 findings, 13 BLOCKING, all accepted; the arms were rebuilt from scratch before dispatch and that rebuild is why the run identifies anything. Translation limb: «باب الناسك والضيف» rendered whole under the new R34 ornamentalist regime, with the first census of where an ornamentalist's hand lands — all ten of the source's figures answered, twelve further sites invented, +6.1% in length.

• v0.2.12 (2026-08-15, S192), ARM-footing-price step 2, RS-20260815c-footing-direction. §10.5 gains items 6a and 6b and §10.6's direction question is discharged in the negative. The run's own instrument gate failed and every registered primary is withheld; what stands is the gate's own measurement — narrated behaviour beats wording on the direction of a perceived social grading, and wording survives on intensity — and two post-hoc figures registered after the failure and before the data existed: four hands of one scene put different people on top at 9 of 10 segments, and two texts identical but for 70 words of marking part company at 8 of 10. The new regime R33 footing-directional is minted, executed and shown not to work: it produces a smoother wrong grading, and none of the source's three direction reversals reaches a judge in any hand. 138 bodies, $0.542027, verifier 219 checks, 0 failures, 4 of 4 mutations caught. Pre-run critic NEEDS-AMENDMENT, 3 BLOCKING of 7 findings, all 7 accepted, and the registered primary replaced before dispatch.

• v0.2.11 (2026-08-16, S191), ARM-answering-figure step 1. §7.12 is new and takes the §7.7/§7.8 unlicensed-mark diagnostic from punctuation and emphasis to sound: RS-20260816-answering-figure, 372 blind bodies over three arms of one whole chapter of «كليلة ودمنة», $0.975246, verifier 1,286 checks, 0 failures, 3 of 3 mutations caught, all four third-party instrument controls unanimous. The registered primary is WITHHELD by its own two manipulation gates and the gates are the result: a sound device that preserves the sentence's meaning was heard at 4 of 13 places where the Arabic has a figure and 0 of 13 where it has none (exact one-sided P = 0.0478), the four audible ones all being multi-member parallel structures. The framework therefore declines to say "compensate" and declines to say the opposite; what it adds to the diagnostic is count the source's matched members, not its sounds. The reading behind that — that the parallel structure crossing for free is what a reader attributes to the author — is marked untested, is post hoc, and is equally consistent with the lead's decoy devices being worse than its compensations, which the run's own failed arm-equivalence gate cannot rule out. Two published Victorian hands separate the same source's sound loci from its plain ones at 6 of 8 against 0 of 8 on third-party English (P = 0.0035), declared secondary before the run.

• v0.2.10 (2026-08-15, S188), ARM-elevation-resolution step 2, and the arm closes resolved at 2 of 2. §7.11 is new and gives Q-e its first measured statement about elevation: RS-20260815-register-room, 36 of 36 blind bodies over three complete lead renderings of Chekhov's «Злоумышленник», $0.321811, verifier 1,475 checks, 0 failures, 3 mutation tests, 3 caught, pre-run critic NEEDS-AMENDMENT with 2 BLOCKING, all 9 findings accepted and two of them implemented in code before dispatch. Substantive changes: §7.11 records that a one-step register policy IS read as raising the source at its low places (+0.50 long, +0.75 short against an unruled −0.25), which refutes the sentence RS-20260814d proposed, and that the effect is larger at one-word items, which refutes the rival explanation as well; that its audible carriers are grammatical — a single corrected pronoun case worth a full point, 49 contractions removed — and that a rewrite-fraction account fails (ρ = +0.345, tested post hoc); and that the frozen translator's log and the blind seats name the same devices. §8 gains a new "not evidenced for" entry: two pairs, two authors, opposite results, so v0.2 carries no general claim about what a moderate register policy does in either direction, and still nothing about where published hands sit. The same-text noise control read 0.0000 over 6 cells; G1 passed at ceiling and licenses detection only, with the interior reading resting on three distinct arm distributions instead.

• v0.2.9 (2026-08-14, S187), ARM-footing-price step 1. §10.5 and §10.6 rewritten on RS-20260814h-footing-price — 72 blind bodies over four arms and six matched segments, $0.378856800, verifier 66 checks, 0 failures, 7 mutation tests, pre-run critic NEEDS REDESIGN with 2 BLOCKING, 18 of 20 findings accepted before dispatch and the run's primary moved as a result. Substantive changes: §10.5 item 2a decomposes the price and takes 70% of it away from the footing — the bulk-matched decoy loses 0.972 of the 1.389 naturalness points, so the registered deference-specific margin (0.417) misses its 0.50 bar; item 2b records the decoy scoring below the plain baseline on social marking, which is RS-20260809h's CLUNKY result failing to reproduce on this item; item 6 is new and is the first statement §10 has about what the marking delivers — a perceived relative grading, in 18 of 18 cells, whose direction follows local device density rather than the source; item 7 is new, post hoc, and names a channel §10 was written as though did not exist, the social relation arriving through ordinary rank nouns in the untouched baseline. §10.6's first open question is discharged and replaced by three successors, one of which is an explicit unresolved: the middle rendering's twelve-token purchase was withheld by the run's own same-text noise control, not reported as a null. A third regime, R28 footing-selective, and a third rendering, T-genji-yomogiu-R28-v1, were minted for the run. The word reader is struck from every claim in §10.5: the judges are three models, Tier D is NOT PASSED, and the subsection stays untested.

• v0.2.8 (2026-08-14, S186), ARM-honorific-hands step 2, and the arm closes resolved at 2 of 2. §10 rewritten whole, on RS-20260814b-honorific-hands (run at S181, $1.109045850, 69 verifier checks, 0 failures, 4 of 4 mutations caught, pre-run critic NEEDS-AMENDMENT with 4 BLOCKING all accepted) and on this session's R06/R27 pair. $0.00 spent this session. Substantive changes, not editorial: §10.3's rank-noun sentence is deleted rather than struck — two 2026 machine hands use the device zero times on a second Japanese source, the whole rank-noun set occurs once in 5,361 words across five hands, and what §10.3 had observed was that 「煙管」 contains a daimyō. The device that carries a grammar-only honorific in a published human hand is a deferential auxiliary verb (Waley's deign), once. §10.1 gains two strata lost at the floor — addressee-polite はべり 0.000 over 19, the prefix 御 0.000 over 14 — and loses its claim that nothing can be recovered. §10.2 gains a fourth failed predictor, G5: a mark riding on a part of speech English also has is not thereby carried. §10.5's materials question is discharged and replaced by the first measurement of what a hand that tries can reach: 52 mechanically extracted sites, 51 carried, 0 with no device available, 15 free and 36 at a named price, +21.1% length (T-genji-yomogiu-R06-v2 / T-genji-yomogiu-R27-v1, regime R27 footing-max, minted here). Marked untested — one hypothesis-aware hand, self-report, judged by nobody. §10.6 is new and carries the two designs this section now needs. Translation limb: 『源氏物語』 ch. 15 「蓬生」 §3-4, rendered twice, 1,227 characters → 836 and 1,012 English words.

• v0.2.7 (2026-08-13, S176), ARM-footing step 2, and the arm closes resolved at 2 of 2. §10 added. RS-20260813e-slot-typology-ja / E-20260813e, $0.502906385 of a declared $0.60; verifier 25 checks, 0 failures, 3 mutation tests, 3 caught; pre-run critic NEEDS-REDESIGN, five findings, three BLOCKING, all five accepted — one of them replaced the design's discriminating cell with a better one. Translation limb: Akutagawa's 「煙管」 completed (T-kiseru-R04-v2, 2,442 words). S1 gains its loss half on a third language family — the pronoun tier reaches English at 0.022 — and gains NO predictor: the positional account offered by RS-20260812i was registered as a prediction and FAILED at 0.200 against a bar of 0.50, with the utterance-final vocative used once in 1,762 words of blind speech. The insertion claim of RS-20260812h is withdrawn on two failures to replicate. Waste $0.061937160, 12.3% — a deepseek/deepseek-v4-pro body that spent its whole cap on hidden reasoning and returned null with effort: low set, and two gemini-3.6-flash cap truncations.

• v0.2.6 (2026-08-11, S156), ARM-idiom-reach step 2, and the arm closes resolved at 2 of 2 on its RESPECIFIED step, not on either criterion it was constituted with. RS-20260811-floor / E-20260811-floor; $0.329161470 against a declared $0.90; 18 of 18 judge bodies, 222 of 222 cells; verifier 73 checks, 0 failures, 5 of 5 mutation tests caught; waste $0.150472500, 46% of the spend, nine dead bodies on one slug's hidden reasoning — note (bmb). Pre-run critic NEEDS-REDESIGN, two passes, 22 findings — 19 BLOCKING — 12 accepted in full, 4 accepted in part, 5 overruled in writing and one REFUTED MECHANICALLY (its headline finding, that a regex matched arm, is false and a regression test says so). §7.3's sentence "the placeless baseline is not placeless" is WITHDRAWN: R22 carries one orthographic nation-mark in 1,969 words, below every one of twelve published texts and 7.7× below its own translator's unruled English, and that one mark is apologise. P1 fails — the median first mark in published English narration is at word 626.5, not 200 — so a page can be unplaced, a book cannot. NAT1 fails at 0.480 CANNOT TELL and M5 at 0.283: what judges quote is mostly not language. SWAP passes on all three seats (0.813 / 0.938 / 0.938) and is clean on two, against an unchanged-duplicate baseline the critic forced in. §7 gains §7.4 and its first measured denominator. R1's text, S1, and every recommendation are unchanged.

• v0.2.5 (2026-08-10, S155), ARM-idiom-reach step 1. The arm stays active at 1 of 2 and neither of its closing criteria fired — the primary is withheld, so it cannot state the register comparison (a) and cannot state a measured refusal (b); step 2 is respecified as the floor measurement §7.3 names. RS-20260810z-idiom-reach / E-20260810z-idiom-reach; $0.172620430 against a declared $1.75, key reconciliation exact to 1e-15; 840 of 840 rating cells; verifier 91 checks, 0 failures, 5 of 5 mutation tests caught; waste $0.00 billed, one dead body at $0. Pre-run critic NEEDS-REDESIGN, 16 findings — 7 BLOCKING — 12 accepted, 4 accepted-in-part, 5 individual remedies overruled in writing; its BLOCKING 1 added the contemporaneous source-first −I arm and moved the primary by 0.167, more than the effect under test, and its BLOCKING 5/6/14 replaced a self-fulfilling positive control with the lead's own frozen −I/+I pair — which is what caught the finding. P1 fails and its null is WITHHELD by F4; P1b reproduces at 0.0000 twice; P3 reproduces a third time at 0.5500 / 0.5833. §7 gains §7.2's successor as §7.3: the placeless baseline is not placeless, and §8's Q-e entry is amended to say so. R1's text, S1, and every recommendation are unchanged.

• v0.2.4 (2026-08-10, S150), ARM-register-reach steps 1–2, and the arm closes resolved at 1 of 2. RS-20260810c-register-reach / E-20260810c-register-reach; $0.510728020 against a declared $2.40, key reconciliation exact to 1e-15; 720 of 720 rating cells; verifier 43 checks, 0 failures, 3 of 3 mutation tests caught; waste 19.3%, four dead bodies, three of them hidden-reasoning truncations on three new slugs at once. Pre-run critic NEEDS-REDESIGN, 20 findings — 4 BLOCKING, 11 SERIOUS, 5 MINOR — 17 accepted, 3 accepted-in-part, 6 individual remedies overruled in writing; its BLOCKING 1 turned the estimand from a device effect into a permission-policy effect and its BLOCKING 3 struck the word exact from the test. §7 gains §7.1's successor as §7.2: the power problem was removed and the comparison was withheld anyway, because the located-idiom permission was exercised at 4 of 60 hand-sites against the respelling's 32 of 60. P3's failure reproduces at loc(B) 0.5833 / 0.5667. R1's text, S1, and every recommendation are unchanged.

• v0.2.3 (2026-08-09, S145), ARM-register-devices step 1–2. RS-20260809g-device-cross / E-20260809g-device-cross; $0.406281508 against a declared $1.80, key reconciliation with a named $0.0597891 gap that is a billed request whose body a shell timeout killed; verifier 134 checks, 0 failures, 3 of 3 mutation tests caught; waste 29.4%, both bodies the same model on Japanese. Pre-run critic NEEDS-REDESIGN, 24 findings, 22 accepted and 2 overruled — its BLOCKING 1+2+6 split the primary into a policy contrast and a conditional contrast, because the two devices were never measured on the same units. §7's design was run: the devices separate, P1a is WITHHELD in both language cells on a power gate that was not moved, and P3 FAILS at 0.5714 / 0.5357 against 0.25 — nonstandard spelling is not placeless and is placed more often than located idiom in all four judge × cell blocks. §7 gains §7.1; its refusal stands with a new reason. R1's text, S1, and every recommendation are unchanged.

• v0.2.2 (2026-08-09, S140), ARM-trajectory step 2. RS-20260809a-trajectory-ja / E-20260809a-trajectory-ja; $0.686836 against a declared $1.60, key reconciliation to 3e-9; verifier 64 checks, 0 failures; waste 7.2% against the previous session's 42.5%. Two pre-run critics on two labs, both NEEDS-REDESIGN, 20 findings, twelve amendments — one of which (A7) rebuilt the positive control after the frozen one collapsed on contact with the material, and one of which (A4) turned the residual check into a positive check that then caught a real classifier defect in the run's own pool. §6 gains Q-d's answer in one language pair and the three qualifications that travel with it; §8's "never been measured" line is replaced. S1, R1's text, and every recommendation are unchanged.

• v0.2.17 (2026-08-16, S202), ARM-invented-ornament step 2, and the arm closes resolved at 2 of 2 with its Done when met and its constituting question unanswered. RS-20260816h-target-set / E-20260816h-target-set; $0.352836 against a declared $0.42; 93 bodies, 4 dead (finish_reason: length, reported, never re-rolled); verifier 323 checks, 0 failures, 5 mutation tests, 5 caught. Pre-run critic NEEDS REDESIGN, 11 findings, 3 BLOCKING; v1 never dispatched, six amendments implementing nine findings, two overruled in writing — its BLOCKING 2 changed the grain to one locus per body, and its MAJOR 7 added the PARALLEL label that turned out to carry the cleanest result on the page. The critic also read the frozen figure inventory and found three misclassifications in it, two of them real; the translation page carries the erratum and its headline counts moved from 20/13/27/6 to 18/11/24/5. The registered primary FAILS at +0.2778 against +0.40 and P1 and P2 fail in the opposite direction, while the level clause holds at 1.0000 / 0.0000. §7.17 is new: it strikes exclusion (a) of this project's standing figure-inventory rule on 9 of 9 bodies, tells a translator to answer the union of a form rule and a sound rule, and adds an untested item on repetition and distance. §7.15 is narrowed in place. R1's text is unchanged and no recommendation is withdrawn.

• v0.2.18 (2026-08-16, S203), ARM-published-figure step 1 — the arm's first step and the first Tier 1 anchor this project has with Arabic as source, A-knatchbull-kalila. RS-20260816j-published-figure; $0 — every limb is reading and writing. Translation limb: «باب القرد والغيلم» whole, the fifth Kalīla wa-Dimna chapter, 1,201 Arabic → 2,220 English under the new R38 repaired-inventory regime, the first rendering made under a rule this handbook recommends rather than under an arm of a manipulation; copy-text collated whole against Būlāq 1937 with eight substantive adoptions, four of them restoring lost narrative and one a preposition that reassigns the fable's cleverest speech. Fifty-four figure loci frozen before any English existed; answering all of them cost 4.4% in length against 4.5% measured independently on the previous chapter. Study limb: Knatchbull 1819 read whole at 72 codable loci across two chapters — ANSWERED 11 (15%), and SHAPE 0 of 18 — with his 1818 preface predicting exactly that, and with §7.17 instruction 1 corroborated by a non-model hand at 3 of 11, not 2 of 2. §7.18 is new and §7.17 instruction 2 is amended in place to say that a practitioner following it does something published practice does not do. No recommendation is withdrawn; all 72 calls are one unchecked hand's and are internal-judgment-only.
• v0.2 (2026-08-08, S133), ARM-marking-work step 2, and the arm closes resolved at 2 of 2 by its second completion condition — shown not to be re-wordable, and closed with the reason. RS-20260808b-discordance-fails / E-20260808b-discordant-marking; $0.851843373 against a declared $1.60, key reconciliation exact to 1e-9; 346 of 348 cells (99.43%); verifier 143 checks, 0 failures, 5 mutation tests, 5 caught. Pre-run critic NEEDS-AMENDMENT, 2 BLOCKING and 6 ADVISORY, all eight accepted, none overruled — its BLOCKING 1 removed the lead's discretion over where the two generated arms' words begin and end, which would have voided P2. The contamination gate demoted the lead out of every primary before dispatch: its close rendering shares 27 contiguous tokens with Garnett on one of the two stories, against 16 for two genuinely independent published hands. F1 fired at 1 discordant site against a floor of 6 and the primary is WITHHELD; the floor was not moved. The finding is that the admission condition did not reproduce — Fleiss κ 0.0584 on the concordant/discordant call against RS-20260807d's 0.786, with the markedness half of the same call from the same seats reproducing at 0.7214. §3 prediction 1 is RETIRED and S1 replaces it. §4 gains the clitic-against-noun-phrase contrast. §7's RU→EN row is strengthened. §8 Q-b gains its third and fourth measurements and changes shape. §8 Q-d is opened. R1's text is unchanged and no recommendation is added or withdrawn.
• v0.2.1 (2026-08-08, S137), ARM-low-pole step 2 and the arm closes. RS-20260808f-placeless / E-20260808f-placeless; $0.441917740 against a declared $0.96, key reconciliation exact to the cent-millionth, 197 of 197 stage-A items and 120 of 120 stage-B cells, verifier 0 disagreements. Pre-run critic NEEDS-REDESIGN, 3 BLOCKING and 7 ADVISORY, all ten accepted, none overruled; its BLOCKING 3 replaced the gate's checker and its ADVISORY 8 removed the last degree of freedom the design had. The repaired gate fired again, on one site in forty-seven, and P1/P2/P3 stay WITHHELD; the gate was not weakened. §8 gains Q-e, which states the register-carriage problem and refuses the recommendation. R1's text is unchanged and no recommendation is added or withdrawn.
• For the whole of v0.1's changelog — the release itself (S091), the stress test (S101), Q-c answered (S106), the obstacle relocated (S116), the census built (S121–S122) — see framework/v0.1/README.md §9. It is not restated here.