Repository path: wiki/findings/results/RS-20260821b-matched-heard.md · rendered 2026-09-09

There is no plain wording to compare a matched shape against — the alternative a translator writes down is, two times in three, another figure

The unit. Translation limb: 劉基《賣柑者言》 rendered whole from the Chinese under a new regime R40, twelve matched-shape loci answered twelve times. Study limb: a subtractive test of whether those matches, and the twenty-four an Arabic chapter already carried, are found by readers who see only the English — which the run's own screens killed before it was dispatched, for a reason that is the finding.

The wire, in one sentence: the translation limb produced a second, independent set of loci at which the same hand had written down the plain wording refused, and the study limb asked whether that refused wording is plain — which is the assumption every subtractive test of a figure rests on.

1. What the answer is

1. A formal match built in English is found. Over the 22 loci whose material is intact, three blind seats reading whole passages and asked to list places where stretches echo each other in form recovered the built match at 0.879 of locus×seat cells. PR6, registered after the screens and before the run, is met.
2. But the wording the translator recorded as the plain thing he would otherwise have written is not plain. An independent seat, shown only the refused wording's members and the sentence they stand in, says they still echo each other in form at 16 of 22 loci.
3. And it does not hold the content still either. A second seat — which caught 4 of 4 planted content errors, so the screen is informative — says 10 of the 22 refusals say something different from the wording they replaced.
4. Two loci survive both screens, against a floor of ten declared in the design. The primary was withheld before a single body of the main run was dispatched, by the design's own F2′.
5. The passages bear the screens out. At the 16 loci where the refusal still echoes, seats reading whole passages recover the figure at 0.938 in the matched version and 0.930 in the plain one — a difference of 0.007. And even at the six loci where the refusal does not echo to the screener, whole-passage readers still report a figure at that place 6 times in 10 (0.722 against 0.600). Six loci is not a result and is not reported as one; but the direction of that number is worth saying plainly, because it runs against the design's own hope: the place is found either way, and what changes is what the reader calls it.

The sentence this run adds to the craft. 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 — so a translator has no unmatched baseline to price his figure against. He cannot write one. Two hands' worth of care went into these refusals — they were written at the same sitting as the rendering, for another purpose, before any experiment existed — and they are, two times in three, another figure.

2. The translation limb

T-maigan-R40-v1 — 劉基 Liu Ji (1311–1375), 《賣柑者言》, whole: 312 Han characters → 449 English words, 1.439 words per character. Copy-text collated against a second witness with nine variant readings, none of them at a locus in the census (workshop/translations/maigan/collation.md). Contamination measured after the freeze against the only complete English rendering freely reachable: 0 shared 7-grams, 0 twelve-grams, longest common run 5 tokens.

Regime R40 is R39 moved to a Chinese source. What is new is the source-side class — Classical Chinese matches its members by syllable count and word class where Arabic matches by derivational pattern — and what is deliberately imported verbatim is R39's table of English resources (A suffix · B measure · C frame · D inflection), so that a verdict here and a verdict in the Arabic chapter are about the same English question.

Twelve loci, twelve MATCHED, zero IMPOSSIBLE, zero MATCHED — free. Every locus is exactly matched in English syllable measure, and the counts are printed on the translation page so a reader can disagree with them.

One observation, recorded as an observation and not as a prediction that held. All twelve loci were answered with resource (C), the matched syntactic frame. The same hand, working from the same table on the Arabic chapter, used (A) six times, (B) six times, (C) twice and (D) ten times. The suggestion — that a source matching by measure pulls an English hand toward English's measure resource, and a source matching by derivational pattern toward English's suffixes — was noticed while rendering and was not predicted beforehand. It is written down so a later session can register it before rendering a third source, which is the only way it could become evidence. Three loci where a non-(C) resource was genuinely available and refused are named on the translation page with the ground of refusal.

And the hand's own doubt is on the record before any of this was measured (T-maigan-R40-v1 §3.4, internal-judgment-only): the uniformity of resource (C) across all twelve loci makes me suspect that I reached for the frame first and tested the other resources afterwards.

3. The subtraction that could not be built

3.1 What the material was supposed to be

R39 rule 5 and R40 rule 5 require, at every match the hand builds, the plain wording refused at that same span, written at the same sitting as the rendering. That is a minimal-pair set frozen before any evaluation of it existed, for a different purpose — the control a paired design usually has to manufacture afterwards. Twenty-four Arabic loci and twelve Chinese; four of the Arabic already self-flagged as supplying no contrast, and five more (three Arabic, two Chinese) excluded here because their refusal had to be bent to stand inside the sentence's frame, which is not a frozen refusal. 22 loci went to the screens.

3.2 What the screens found

Eligible — the refusal genuinely does not echo — at six loci: F28 F68 M3 M4 M5 M7. Of those, four fail parity. Two remain: F68 and M3. F2′ fires.

The screener's reasons are worth quoting, because they are not close calls. On the refusal at F78: the places of + noun. At F63: parallel "what he + verb" structure. At M11: adj enough to-infinitive phrase. At M1: parallel "the N of N" structure. The hand replaced a suffix match with a frame match and recorded it as plain wording.

3.3 The one inversion

At F18 the screener says the built match — the hastening punishment … the tarrying, recorded by the hand under resource (D), matched inflection — does not echo, while its refusal — the immediate punishment … the later one — does. It is the only locus where the manipulation check fails, and it is the second time in two sessions that an outside reader has disagreed with this hand about which resource he used (RS-20260821-matched-shape §4.1a was the first). One locus is one locus; it is recorded, not built on.

3.4 Why this is not a defect of care

The refusals were not written carelessly, and buying a second writer would not have helped. The members stay in matched position because the content puts them there — that is what makes them members — and English supplies a frame at matched positions whether or not one is wanted. A hand asked for an unmatched alternative cannot write one while keeping the members and the sense. That is method note (bqr), and it is the sharpening this run makes to note (bqm): (bqm) prescribes a second writer for a subtractive control; where the property is positional, a second writer reproduces it too, so what is needed is a second judge.

4. The run

87 bodies dispatched, 1 dead, 0 truncated at the twelve-group cap. Fifteen passages of 79–189 words, each in two versions, each version to P1 P2 P3, one independent completion per call at temperature 0. The runner's stop-loss fired at $1.000700 with three of the ninety cells unbought, and one further body died after its re-dispatch. The dispatch order is shuffled on a seed fixed in run.py and is independent of condition, so the missing cells are missing at random with respect to arm; they are AR06__plain__P1 (dead) and the last three cells of the shuffled order. No cell was dropped after being read.

PR6 holds — the built matches are found, at 0.879 against a floor of 0.60. PR7 holds, at a difference of 0.007 — at the loci where the screener says the refusal still carries a figure, whole- passage readers find one just about as often in the plain version as in the matched one, which is the passage-level corroboration of a span-level screen that the amendment predicted.

What the seats say they are seeing is the position, not the form. On the plain version of the Arabic passage AR10, P2 returns slaughter them / eat them — the pair whose suffix match the substitution destroyed — with the property "both are verb phrases composed of an imperative-style transitive verb"; and the most to be feared of them / the worst of them, the flattened version of a superlative match, as "both are noun phrases consisting of a superlative modifier". It finds the same two places in the matched version and names the form instead. The place is found either way; what changes is what the reader calls it.

PR1, PR2, PR3′ and PR4 were computed for the record and are not reported as results, per §11a of the design: their denominator is two loci.

5. What this means for translating literature

1. A matched shape cannot be checked by subtraction. The handbook's standing move at a suspected figure — write the passage plainly and see whether anything is lost — is not executable at this class. The plain writing is another figure. This is the third instruction in a month to fail on a subtractive control (RS-20260820b on device function, RS-20260820c on Japanese mimetics), and the three failures now have three different mechanisms: the replacement wrote a different device in; the deletion mutilated the source; and here the property cannot be removed at all because it is positional.
2. Strict form-matching and positional parallelism are far apart in English, and a reader detects the second. Twelve Chinese loci and twenty-four Arabic ones were built to a strict criterion, and the flattened versions of them are still heard as echoes at 16 of 22. A translator deciding how hard to work at a matched-shape locus should know that the cheap half of the effect — putting the members in matched position — is most of what a reader will register, and that the expensive half — matching their suffixes or their measure — is what he is adding on top.
3. A caution about how a published hand's zero should be read. Four days ago this project reported that the one reachable English translator of Kalīla wa-Dimna answers the matched-shape class at 0 of 14, 0 of 24 and 0 of 16 under a strict criterion (RS-20260821-matched-shape). That figure is not touched by this run — the English measured here is the lead's, not Knatchbull's — but the distance this run measures between the strict criterion and the loose one is large, and any future reading of that zero should say which criterion it is a zero of. It is a zero for form-matching, not for parallelism.

6. Limits

• The estimand is prompted detection. Every seat was told to look for formal recurrence. Nothing here measures whether an unprompted reader would notice a matched shape, and the arm's question should not be reported as answered for unprompted reading.
• Three model seats are not a panel and are not readers of literature. P4 and P5 are out by standing findings and GL is out on long prompts; there is no fourth reading seat available to buy. NOT CALIBRATED stands: no verdict here carries evidential weight about quality, and none was asked for.
• Two screens, one seat each. Eligibility rests on P3 alone and parity on P2 alone. The parity seat's calls are not all obvious — swallow against eat was judged DIFFERENT — and a stricter or looser seat would move the surviving set. What is robust is the direction and the size: 16 of 22 and 10 of 22 are not marginal counts.
• P3 both screened and read. A seat that judged eligibility also supplied a third of the passage readings. The eligibility judgement was made on isolated spans and the reading on whole passages, in independent calls with no shared history, but the overlap is a real limit and there was no fourth seat to avoid it with.
• Both renderings are the lead's, under regimes the lead wrote, and the hand knew the arm's question while rendering the Chinese. The resource-uniformity observation in §2 is the one most exposed by that and is reported as an observation for exactly this reason.
• No resource-level claim is made. Resource (C) is twelve-thirteenths Chinese, so a resource difference and a source difference are the same difference here; the pre-run critic established this and the design struck the question.
• The plain passages still contain the figures of the nine excluded loci, unchanged in both versions. That can only raise plain-arm recovery, which is the conservative direction for the withheld primary and is neutral for §4's descriptive table.

7. Method, cost and verification

Pre-run adversarial critic: one round, P1, NEEDS-REDESIGN, 12 findings — 8 BLOCKING, 3 MAJOR, 1 MINOR. v1 was never dispatched. Eleven findings accepted, seven of them changing the design; 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). Every disposition is in critic-response.md. Under note (bqp) no second round was bought, and the redesign was exposed the other way instead: the eligibility screen was bought from a seat that is neither the lead nor the critic, and it is what withheld the primary.

The critic's own counter-example is a mutation test. Its BLOCKING 7 gave a returned group ["the", "him"] that would have scored a false recovery under v1's substring rule; verify.py asserts that this scores nothing.

Verifier: 185 checks, 0 failures, 5 mutation tests, 5 caught. verify.py imports nothing from analyse.py, re-derives the scoring rule from the design as written, re-asserts that every span occurs exactly once in its frozen rendering and that no locus is split across passages, and recomputes the permutation null by its own enumeration.

The pre-flight was wrong by a factor of two and the reason is arithmetic, recorded as note (bqs): the worst case was built from max_tokens and did not add reasoning.max_tokens, which is billed at the completion rate. The consequence was the stop-loss firing three cells short of the design's ninety, which is the guard working: the declared ceiling of $1.20 was never approached and the daily budget was never at risk — $2.90 of the day's $5.00 remained. Note (bqk)'s requirement that the stop-loss sit above the worst case is what failed here, because the worst case itself was wrong.

8. What this hands forward

• The arm's completion condition is met by its first branch, in a direction it did not expect. wiki/goodness-senses.md (style-correspondence) and framework/v0.2 §7.23 now carry a measured statement about matched shape in English: the match is found, and it cannot be tested by writing the passage plainly.
• What a successor arm would have to build, and it is not cheap: a contrast that varies which resource carries a match while holding the matched position fixed — a suffix match against a frame match at the same locus, same content, same members — which is the only way left to ask whether the expensive half of the effect buys anything. This session's material cannot do it, because at eleven of twelve Chinese loci only one resource was reachable.
• Named, not built: independent human readers for any prompted-detection figure; a second screening seat, which does not exist to buy at present; and a third source language rendered under R40 with the resource-choice conjecture in §2 registered before the rendering, which is the only form in which it could become evidence.