Repository path: wiki/findings/theory/TH-20260724-translation-distance-axes.md · rendered 2026-09-09

Two distances, not one — what actually changes when the language pair changes

The project's first theory page (charter §3, poetics track). It is a model, not a finding: an attempt to say what the precedent anchors jointly show, in a form specific enough to be wrong. Every claim below carries its evidence, a confidence, and a falsifier. The synthesis is the lead agent's, over five close readings that are themselves internal-judgment-only; one sub-claim is anchored to Tier 2 and is marked where it appears.

Why this page exists

The project is chartered for "translating literature between languages" with a Japanese→English start (charter §1). Until this session every piece of evidence in the repository came from that one pair. That is a live risk, not a hypothetical one: a typology derived from J→E and a framework whose recommendations are tuned on J→E will encode the properties of that pair as though they were properties of translation. The cheapest correction available is more pairs, chosen so that they differ from J→E in known ways.

Five precedent anchors now exist, built to a design rather than to convenience:

The axes are ordinal judgments, not measurements (§"What this does not establish"). Structural distance means: how much of what the source encodes in grammar — inflectional morphology, grammaticalised social deixis, aspect, gender, animacy, word order, script — has no counterpart category in the target. Cultural-referential distance means: how much of the source's referential world the target's readers already possess. The far/near-reversed cell would need a pair like Chinese→Korean, and is a named gap.

A third column appeared on 2026-07-25 (S018), and it is not a decoration. A-beowulf-ingeld sits in the same two-axis cell as A-shaw-spider-thread — far/far — and behaves differently from it in two large, specific ways (false friends of register exist; some realia copy free). What separates them is neither axis: it is how much of the source's word-stock the target can simply take over. That variable also separates A-yosano-yomogiu from what a zero/far pair "should" do, which is what forced C2's correction in the first place. Lexical inheritability varies independently of both original axes and predicts things neither of them predicts, so the two-axis model is now a three-axis model, and this page's title has outlived it. Renaming is deferred until the third axis has more than the four cells below.

The near/far cell was filled on 2026-07-25 by intralingual translation, and filling it cost the model something. A-yosano-yomogiu reads three renderings of one 951-character passage of the Genji — Yosano Akiko's (1938–39), Shibuya Eiichi's (contemporary), and the lead's own English (T-genji-yomogiu-R04-v1, translated and its log frozen before either Japanese rendering was read). The results corrected C2, extended C3, narrowed C4, and gave C1 a control it had never had; all four revisions are recorded in place below. The single most consequential one: at maximum lexical overlap and enormous cultural distance, the mediation load did not behave as C2 predicts, because the target could simply inherit the source's signifiers.

The model

Translation difficulty is not a scalar. Pairs do not sit on a line from easy to hard. They differ in which goodness senses carry the load, and the two axes above predict which. That is the whole model; the claims below are its content.

C1 — Grammaticalised categories the target lacks are lost, and repaid, if at all, in the lexicon

Confidence: moderate-to-strong for the loss (restored 2026-07-27, S038), moderate for the repair. Downgraded 2026-07-25 from "strong" when the Baudelaire row was retracted; restored when A-sasaki-kuroneko supplied the English-source instance the downgrade was about. ~~Three~~ Two independent instances, ~~three~~ two pairs, two translators, 1922–1930. The third instance was retracted after second-reader verification (RS-20260725-anchor-verification): Poe does not split the two cats by pronoun, so the Baudelaire row was not an instance of a target-lacking grammatical category. Both surviving instances have English as the target.

The regularity is not that hard things are hard. It is more specific and more useful: meaning carried by grammar is not translated, it is transcoded into lexis — and the transcoding costs systematicity. A source category applies automatically, everywhere, at every occurrence; its lexical repayment applies where the translator remembers and where a word happens to be available. What was continuous becomes intermittent. This is why the losses read, in all three anchors, as flattening rather than as error.

This paragraph previously read: "The Baudelaire row does work the other two cannot. The project's evidence base is otherwise all J→E, which could easily encode the assumption that grammar-borne meaning is a property of foreign, distant, or exotic languages. English has it too and loses it to French. The regularity is about category mismatch, not about which language is strange."

It no longer holds, and the concern it answered is live again. The Baudelaire row was the only evidence in this project for a target-lacking category with English as the source, and it was withdrawn on 2026-07-25 after three blind second readers, two adjudicators and a lead re-read agreed that Poe's text does not contain the contrast the anchor described. C1 therefore rests on two instances, both with English as the target (Japanese→English and Russian→English). The claim that "the regularity is about category mismatch, not about which language is strange" is now a conjecture the evidence does not reach — it is what one would expect if C1 is a fact about category mismatch, but nothing here distinguishes it from the alternative that the project has simply only ever looked at translations into English. Filling this gap is the highest-value single addition to C1: one precedent anchor in which English is the source and the target lacks an English grammatical category, in a pair not conflated with lexical overlap. (RS-20260725-anchor-verification; added to wiki/program.md Slate E.)

FILLED 2026-07-27 (S038), by A-sasaki-kuroneko, and C1 survives the test its named gap was for. English → Japanese, on the same source text as the retracted row — Poe's "The Black Cat" — so the project now has one English source into two targets, one at maximal lexical inheritability (French) and one at zero (Japanese). Japanese is the right instrument in a way French is not: French has a third-person pronoun and a gender system, so what it does with Poe's animacy is a choice between categories; Japanese has no obligatory third-person pronoun, no article, and no number on count nouns.

The category was repaid, lexically, at 5 of 15 sites — 33%, and intermittency is the claim. Every masculine third-person token in the story was enumerated and its referent adjudicated: 15 refer to a cat (13 Pluto, 2 the second cat). 佐々木 renders five of them with 彼 — a nineteenth-century calque made for translation out of European languages, marked in a way English he is not — and the other ten with a bare noun or nothing. What is continuous in the source becomes intermittent in the target, with English as the source, in a pair with no lexical overlap to confound it.

And the repayment is placed, not scattered. All five 彼 fall where the cat is an agent — following, being seized, biting, walking about the house — and none where it is an object of violence; Poe himself shifts Pluto to it at the mutilation and through the whole hanging, and 佐々木 shifts with him (「そのかわいそうな動物の咽喉をつかむと…」). The translator tracked the source's use of a category his language does not have, at a third of its sites.

The conjecture the evidence did not reach is now reached, and it holds. "The regularity is about category mismatch, not about which language is strange" was downgraded on 2026-07-25 because both surviving instances had English as the target. It has an instance with English as the source and the strange-language reading is not available for it: English is the source, Japanese the target, and the loss runs source→target exactly as in Shaw and Garnett. C1's confidence on the loss returns to moderate-to-strong: three instances, three pairs, three translators, and for the first time both directions. The repair clause stays moderate — three instances is not many, and the two English-target ones are still 1922–1930.

Revision trigger #2 does not fire. It requires "a target device deployed at every occurrence"; one site in three is the clause holding, not failing. (It also specifies a contemporary translation, and 佐々木 died in 1943.)

A caution the cell adds about its own reading, and it bears on the 2026-07-25 retraction. 佐々木's five 彼 are all Pluto; the second cat gets その動物, その猫, それ, 一匹の畜生, 怪物, そいつ and never 彼. So the Japanese has a personhood asymmetry between the two cats that Poe's English does not have — the very distinction three blind second readers, two adjudicators and a lead re-read agreed was absent from the source. Its second-cat leg is 0 of 2 tokens and licenses nothing by itself; what it does show is that an asymmetry of this kind is something a translation can manufacture out of a category the target does not require. That is a stronger reason to keep the retraction than the retraction gave.

What the cell shows about the other half of C1 — forced explicitation — is a limitation, and it is new. A-beowulf-ingeld and A-yosano-yomogiu found forced explicitation pervasive: a determiner on every count noun, "she / he / her father" at every clause. Working out of English, the lead's frozen log records two sites in 880 words. The asymmetry looks structural rather than accidental: a category the target obligatorily has forces a decision at every occurrence because the slot must be filled, while a category the source has and the target lacks presents no slot — the information does not survive, silently, without asking. So forced explicitation and silent flattening are not two names for one phenomenon: the first generates decisions and shows up in a translator's log, and the second generates none and does not.

Which puts a bias under RS-20260727-log-typology. Twenty of the project's twenty-one frozen logs are of translations into English, and that result counted C4b (determinacy the target compels) at 14 decisions in 11 of 21 logs. On this evidence C4b's rate is partly a fact about English being the target. Recorded as a limitation on that result, not as a correction to it — one passage, one translator.

A third finding, and it is about C2 rather than C1. Japanese has optional devices for both categories it lacks obligatorily — a numeral-classifier phrase for indefiniteness, demonstratives for definiteness — and at the one clean site (It was **a** black cat) 佐々木 takes it (「一匹の黒猫」) and the lead refuses it (「黒猫であった」), the lead's log giving the reason. Both then spend an identical demonstrative budget (34 each) on different distributions. Where the target permits a free option the fork is elective, and two translators elect oppositely — the A-yosano-yomogiu shape, recurring for a grammatical category rather than for realia, which suggests the elective/forced distinction is more general than the case it was found in.

Note the control, from the same texts: what is carried by content ports intact. Chekhov's unaddressable letter — the whole pathos of "Vanka" — survives Garnett unaltered; Akutagawa's symmetrical morning/noon frame survives Shaw; Poe's plot survives Baudelaire. The losses are not distributed by importance. They are distributed by what encodes the meaning.

Falsifier. A precedent anchor in which a target-lacking grammatical category is rendered systematically — by a target device deployed at every occurrence rather than by scattered lexical choices — would break the "costs systematicity" half. (Cohn's 54-occurrence na moshi apparatus, per ES-20260724-fluency-and-foreignness, is the nearest known candidate, and is a lexical item, not a category — so it does not yet count.)

Amended 2026-07-25 (S017), from A-yosano-yomogiu. C1 gains a control and a second half. The intralingual anchor cannot fill C1's named gap — English is still the odd target — but it does something C1 has never had: it shows the same source category (the classical Japanese honorific system, which marks deference obligatorily on every verb and possessive) rendered into a target that has the category and a target that does not, side by side.

• Shibuya (contemporary modern Japanese) retains it systematically, deploying modern keigo at essentially every site the classical marks (「お訪ね申す人はないお身の上である」·「お立ち寄りになる」·「お考えつきにならない」·「お過ごしになっている」). This is what "a target device deployed at every occurrence" looks like — the thing C1 says does not happen when the category is missing, observed happening when it is present. It is a control, not the falsifier: the falsifier requires the target to lack the category, and modern Japanese does not.
• Yosano (1938–39) had the category and largely dropped it, writing much of the narration flat. Availability does not entail retention. C1 is a claim about what a pair makes possible; what a translator does inside that space is governed by purpose. Any framework recommendation that reads C1 as "near pairs preserve grammar-borne meaning" would be wrong.
• The English cannot do either; the layer is simply absent, repaid intermittently by a half-step of formal diction (T-genji-yomogiu-R04-v1 log §2b).

The second half of the cost: forced explicitation. Classical Japanese identifies who is acting through honorific level rather than by naming. Shibuya, having kept the keigo, keeps the ellipsis. Yosano, having dropped it, must name people constantly (「…人をも持たない女王であった」, 「末摘花はそんな趣味も持っていない」). The lead's English is forced the same way at every clause, and its translator's log reached this independently, before Yosano was read. So a lost grammatical category does not only make repayment intermittent — it adds determinacy the source did not have. C1's regularity should be read as: meaning carried by grammar is transcoded into lexis, at a cost in systematicity and at a cost in openness. Two translators, two targets, the same compensation forced by the same loss.

Extended 2026-07-25 (S018), from A-beowulf-ingeld: forced explicitation is not specific to a lost category. The Japanese cell found it as a consequence of a category the target lacks. The Beowulf cell supplies the mirror image — a category the target obligatorily has and the source does not.

• wīf. One Old English word covers woman and woman-as-spouse, and the passage's subject is a woman being converted into an instrument of policy. Modern English forces the split, and at the two sites (mid þȳ wīfe before the marriage, wīf-lufan after it) no translator holds one word across both except Morris, who manages it only by archaising. Gummere writes "the woman" then "wife-love"; the lead writes "this woman" then "his wife"; Kirtlan deletes her — "set at rest by that betrothal" — the most explicit choice of the four, and it removes the person from the clause that is about using her.
• Articles. Old English has no indefinite article and an optional demonstrative; every count noun in all four modern renderings carries a determiner the source did not have to choose. It runs under the whole text rather than at named sites.

So the regularity is about asymmetry in obligatory distinctions, in either direction, not about loss. Where the pair's obligatory categories differ, the translation states things the source left open — whether the mismatch is a category going missing or a category being imposed.

Scope condition added 2026-07-28 (S044), from S-luxun-yingyi — the first time a non-Anglophone primary has bounded a claim on this page rather than illustrating one. C1 says a target-lacking category is transcoded into the lexicon at a cost in systematicity. It presupposes that the target's inventory is fixed, and that presupposition has never been stated because no source in this base contested it. Lu Xun does contest it, in print, as a programme: 「裝進異樣的句法去，古的，外省外府的，外國的，後來便可以據為己有」 — pack in alien syntax, ancient, from other provinces and prefectures, foreign, and afterwards make it one's own — with an explicit schedule (import → digest → part naturalises, part is weeded out), a cited precedent (Europeanised grammar in modern Japanese) and a cited completed instance (罷工, coined 1925, understood within six years).

This does not refute C1 and must not be read as doing so. The two claims are about different objects: C1 is about what happens inside one translation, Lu Xun's is about what happens to a language across many, over decades. But the boundary between them is a real one and it is now on the page: C1 holds where the translator treats the target's inventory as fixed. Where the translator treats it as mutable — and some do, deliberately — the "cost in systematicity" is not a cost but the intended mechanism, and C1 does not describe what is happening. A framework recommendation derived from C1 would be advice to a translator of the first kind and would be simply wrong for one of the second.

What would test it, and the project cannot run it. Whether the import programme actually reshaped modern Chinese syntax is a corpus question over a century of Chinese prose. It is named here as out of reach rather than as an open item, so that no later session mistakes Lu Xun's prediction for the project's evidence.

C2 — Cultural distance, not structural distance, decides how much work cultural-mediation does

Confidence: moderate. Three cells, and the sense's load tracks one axis cleanly.

• Far/far (Shaw). cultural-mediation dominates the anchor: one class of Buddhist-cosmology realia handled four ways in 1,000 words, and the page's headline finding.
• Far/near (Garnett). Most realia are cheap, because the worlds overlap — a cobbler's last is a last, an icon is an "ikon," matins is "the midnight service," kopecks stay kopecks. But a residue does not overlap (a proverb, "Sidor's goat"; a folk custom, going out "with the star"; a snuff-taker's joke), and in the residue Garnett's answer is uniformly copy-opaque: an English surface with nothing under it, exactly like Shaw's 三途の河.
• Near/near (Baudelaire). The sense nearly idles. Monks, hogsheads, gin, a crow-bar, a cellar all pass without mediation, and even the loaded proper name has a ready target form: "Pluto" → "Pluton."

The proper-name slot makes the axis visible in a single variable across all three: Shaw must romanise (三途の河 → "Sanzu-no-Kawa," inert), Garnett must choose (transliterate "Kashtanka" or translate "Eel" — and does both, one in each half of one sentence), Baudelaire has nothing to choose because French owns its own form of the name.

Crucially, the difficulty does not vanish in a near pair; it migrates — to style-correspondence (C3). A framework that read "few realia problems" as "easier pair" would be reading the wrong instrument.

Falsifier. A structurally far, culturally near pair whose realia turn out to be as costly as J→E's — or a near/near pair with a heavy realia load — would break the mapping. The obvious test is a culturally distant but structurally near pair, which the three anchors do not include.

CORRECTED 2026-07-25 (S017). The obvious test was run, and C2 failed it — in the direction the falsifier's wording did not anticipate. A-yosano-yomogiu is the culturally distant, structurally near case in its purest available form. C2 predicts a heavy cultural-mediation load. The load was light for both intralingual translators, and not because the items are easy: a modern Japanese reader no more knows what 紙屋紙 is than an English one does.

The reason is that Yosano and Shibuya are not obliged to decide. 浅茅 · 蓬 · 葎 · 寝殿 · 野分 · 禅師の君 · 紙屋紙 · 数珠 — all can simply stand, at no cost, and the opacity that results is inherited: it is the same opacity a Heian reader unfamiliar with the imperial paper works would have met. When the lead's English writes "Kamiya paper" it manufactures a new opacity, a foreign string, which the source did not contain. Copy-opaque handling across languages always creates foreignness; intralingual copy never does.

The corrected claim:

cultural-mediation load is a function of cultural-referential distance × whether the target's script and lexicon permit the source's own signifiers to be inherited. Where inheritance is free, mediation stops being a forced fork and becomes an elective one.

"Elective" is not a hedge; it is what the anchor observes. Given the identical free option, Yosano substitutes at five sites (陸奥紙→檀紙, 御厨子→書物棚, 総角→牧童, 下衆→下男, 受領→地方官) and Shibuya at one. Two translators of one language, offered the same costless copy, choose oppositely. In the three interlingual anchors that divergence is not even possible, because there is no free copy to diverge about — which is why three cells of interlingual evidence could not have surfaced this.

Confidence: C2's mapping is downgraded from "moderate" to "holds for interlingual pairs; superseded by the inheritability formulation in general." The original claim is not deleted — across the three interlingual cells it still describes the data — but it was reading a proxy. Cultural distance predicted mediation load in those cells because, between languages, cultural distance and non-inheritability move together. Intralingual translation pulls them apart, and it is inheritability that tracks the load.

New falsifier for the corrected claim. A pair with free inheritance of signifiers (a shared script and a large shared lexicon — Chinese→Japanese kanbun, or classical→modern Chinese) in which realia handling is nevertheless a heavy forced fork; or an intralingual case where translators converge on handling rather than diverging.

TESTED AND SURVIVED, 2026-07-25 (S018), by A-beowulf-ingeld — and repaired on one point.

The corrected C2 makes a prediction the original C2 does not: a pair that is intralingual but not inheritance-rich should behave like an interlingual pair, because it is inheritability and not intralinguality that governs the fork. Old English → modern English is that pair — same language by name, continuous transmission, inherited script, but a lexicon that mostly did not survive. It behaved as the corrected claim predicts. Ten culture-bound items in 56 lines are forced forks and all four translators forked on all ten (flet, bēah-wriða, ealu-wǣge, lind-plega, here-grīma, wīg-bealu, bill, māððum, duguð, Wiðergyld) — a heavier mediation load than A-garnett-vanka carries across a whole story. The rival reading of the Japanese result — that intralinguality as such makes mediation elective — is refuted.

The free-copy residue is small and its shape is informative: medu, ealu, bēor, gold, īren, land, word, blōd, fæder, sunu, dohtor. Substances, drinks and kinship copy free; institutions and material culture fork. What a thousand years destroyed was not the language but the society the words named, and the mediation load tracks the society.

The repair. A-yosano-yomogiu concluded that copy across languages always manufactures a new opacity while intralingual copy never does, because the resulting opacity is inherited. That is too strong. Old English proper names are transparent common nouns — Wiðer-gyld is literally requital, Frēa-waru is lord-protection, Heaðo-beardan is battle-beards — and all four translators copy them and all four spend the transparency, leaving strings as opaque as Shaw's "Sanzu-no-Kawa". Only Kirtlan pays anything back, with a footnote. So:

Copy manufactures opacity wherever the target's readers no longer hold the sense the signifier had. The language boundary is not the variable; reader distance is. In the Japanese cell the two coincided — a modern reader holds the classical word's sense and merely lacks the referent. Here they come apart.

This makes C2 and the pair-indifference observation at C3 two sides of one thing: part of what the project has been scoring as translation difficulty is set by how far the reader stands from the source's world, and sharing a language buys no discount on it.

C3 — High lexical overlap trades cultural-mediation cost for style-correspondence cost, and can foreclose the foreignisation choice

Confidence: weak-to-moderate — one anchor, one pair. Stated fully anyway, because it is the most consequential claim here and the most falsifiable.

The near/near anchor shows two effects that only a shared word-stock can produce.

(a) The cognate is a false friend of register. Poe's style is high-Latinate diction against a Germanic base; the contrast is the style. Baudelaire's cognates — décharger, élucider, fantôme, opiniâtreté, indicible — are each defensible and each flatter, because a Latin root that marks elevation in English is ordinary stock in French. Maximum lexical fidelity, minimum stylistic marking. Closeness causes the loss: in a far pair no cognate exists to tempt anyone, so register must be chosen consciously.

(b) The pair can delete the foreignisation option. Poe italicises _barroques_ to make a French-flavoured word ring foreign inside English. Baudelaire keeps the italics; in French the word is unremarkable, so the marker now marks nothing — form preserved, function evacuated. The project has been treating domestication↔foreignisation as a choice whose cost is diagnosable ("what did the fluency cost?", ES-20260724-fluency-and-foreignness). Here there was no choice to make. Domestication can be structurally imposed by the pair rather than adopted by the translator — which means the essay's diagnostic needs a prior question: was there a fluency choice available at all?

And where the pair erases foreign residue, a translator who wants stylistic marking back must manufacture it. Baudelaire does so two ways: re-foreignisation (leaving "Gentlemen" in English inside the French, three times) — the exact mirror of Shaw romanising 三途の河 — and coinage (hyperdiabolique, antihumain, intraduisible where Poe wrote "unutterable"). The coinage habit is Tier-2 anchored: Wallaert (2012) documents it across the Poe corpus and reads it as "an appropriation of the source text… hidden behind a smokescreen of apparent literalness" (S-wallaert-baudelaire-neologisms). So the near pair's repair is real, and it is also a re-authoring.

Falsifier. A structurally near but lexically non-overlapping pair (the anchors conflate these two variables — English↔French is an extreme of borrowing) that shows the same register flattening would show overlap is not the mechanism. Conversely, a near/near translation that preserves marked register without coining would weaken (b).

EXTENDED 2026-07-25 (S017), from A-yosano-yomogiu. C3 is the claim the intralingual cell strengthens — and it generalises past "cognate".

(a) becomes: any formally available but semantically drifted form is a false friend of register, and time produces drift as reliably as borrowing does. The anchor's headline case is one word doing structural work. うるはし occurs three times in 951 characters — of the house, of her paper, of her person — asserting that all three are the same quality: formally correct, well-ordered, and not alive. Modern Japanese has the word (麗しい) and its sense has drifted to beautiful, lovely. Both intralingual translators saw the shared form and both declined it; having declined it, each fell back on local paraphrase, and local paraphrases at three sites do not coincide:

	Yosano 1938–39	Shibuya	lead (English)
thread held across the three nodes	0/3	2/3 (「きちんとした」 ×2)	3/3 ("impeccable" ×3)

The translator at zero lexical overlap held the thread; the translator at maximum overlap dissolved it entirely. Overlap did not merely fail to help — it hurt, by supplying a candidate that had to be refused independently at each site, which is exactly the condition under which a thread breaks. The lead, with no shared word to refuse, went looking for one English word that would survive all three positions and found one (log §2a records the search and its cost).

The same shape recurs on the aesthetic-evaluative vocabulary, which no modern target preserves. 「めざましき」 names something as intolerable given who is doing it — its meaning contains a rank differential — and modern 目覚ましい has drifted to unambiguously positive. All three translators substitute a plain modern moral judgment (無礼 · けしからぬ · "effrontery"). あはれにいみじきこと多かり goes the same way in all three. The source's aesthetic lexicon is lost at maximum lexical overlap exactly as it is at zero, which is the clearest single demonstration available that lexical overlap is not the operative variable — semantic distance is, and lexical overlap only determines whether a tempting wrong candidate is on offer.

(b) is not extended: the intralingual pair does the opposite of foreclosing the choice. In intralingual translation foreignisation takes the form of retaining the archaic word, and both options are fully live: Shibuya foreignises heavily (禅師の君, 御厨子, 陸奥紙, 下衆, 葎 all retained), Yosano domesticates (書物棚, 檀紙, 下男, 牧童, 地方官). So overlap widens the choice on realia while narrowing it on register-bearing words. C3 should not be stated as "overlap forecloses choice"; the near/near foreclosure in A-baudelaire-chat-noir was about a marker (italics) whose function the target had evacuated, not about overlap as such.

Also observed, and stated as an observation rather than a claim: some problems are pair-indifferent. 「煙絶えて」 — "the smoke ceased," an ellipsis a Heian reader completes without effort — was repaired at the same site by all three translators with the same two-to-four-word scaffold: 「廚の煙が立たないで」 (Yosano) · 「炊事の煙も上らなくなって」 (Shibuya) · "The smoke of the kitchen fires died out" (lead). Three translators, two target languages, ninety years apart, supplying the same missing noun. If this generalises, some part of what the project has been treating as translation difficulty is reader-distance difficulty, on which the intralingual translator gets no discount at all. It is one site in one passage, and the lead's convergence is exposed to its own contamination: high — so it goes in the program as something to test, not into the model.

Confidence: C3 rises from "weak-to-moderate — one anchor, one pair" to "moderate" for (a) as reformulated — two cells now, two mechanisms (borrowing; drift), three translators. (b) stays weak-to-moderate and narrows to markers whose function the target has evacuated.

SHARPENED INTO A MECHANISM, 2026-07-25 (S018), from A-beowulf-ingeld. This is the strongest result the theory has.

The Beowulf cell offers fifteen sites in 56 lines where modern English holds the descendant form of an Old English word, and four translators (Morris 1895, Gummere 1909, Kirtlan 1913, the lead) each deciding at each site whether to take it. Scoring rule and full site list are on the anchor (§4.3); the totals:

drift	sites	Morris	Gummere	Kirtlan	lead	all
total — modern sense has no overlap (drēam→dream, wine→wine, rǣdan→read)	3	0/3	0/3	0/3	0/3	0/12
partial — modern sense wrong but locally survivable (glæd, eorl, wīf, duguð, mōd, folc, bill, hyrde, gladiað, wīf-lufu)	10	9/10	5/10	5/10	1/10	19/40

The within-root control is the cleanest evidence in this project. glæd (2026) and gladiað (2037) are the same root, eleven lines apart, in one passage, translated by the same four people. At 2026 three of four take the reflex — "the glad son of Froda," in those exact words, in Morris, Gummere and Kirtlan. At 2037 none of the four do; all write "glisten" or "gleaming". Nothing differs but this: at 2026 the modern sense is merely wrong (a prince can be pleased, so the phrase passes as English), while at 2037 it is impossible (heirlooms cannot gladden).

C3(a), restated. The false friend of register operates inside a window. Below it, drift is slight and the reflex is simply usable — no problem. Above it, drift is total and nobody is tempted — no problem. The damage is done in the middle band, where the modern form yields a locally grammatical, locally plausible sentence that says the wrong thing. 0/12 against 19/40 is a threshold, not a gradient.

This retro-explains the Japanese cell rather than merely agreeing with it. うるはし→麗しい and めざまし→目覚ましい sit in the window — both modern words are usable adjectives of approval — which is why both intralingual translators had to refuse them site by site and why the thread broke. And it explains why the lead, at zero overlap there, held the thread 3/3: having no reflex at all is the same condition as having a totally drifted one. No candidate, no repeated refusal, no drift in the answer. Lexical overlap hurts a thread only when it lands in the window.

A second mechanism, new to the project: archaism buys inheritability back. Morris takes the reflex at 9/10 partial-drift sites against 4, 5 and 1. By pitching the whole text archaically he licenses his reader to read words in their older senses, and those senses are often the Old English ones — "that wife" for wīf meaning woman; "the herd of the realm" for rīces hyrde; "thou shouldest arede" for rǣdan; "Peace-sib of the folk"; "after bite of the bill". The payoff is not merely lexical: duguð and dugan are one root and the poem plays on it, and Morris is the only one who holds that etymological thread complete — 3/3, against Kirtlan 2/3, Gummere 1/3 and the lead 0/3. The cost is real too — "the herd of the realm" reads, to a modern eye, as livestock.

So on a diachronic pair, foreignisation is not only a stance but a capability: it moves words out of the window by changing what the reader is asked to expect, and it thereby makes reachable a class of source-internal correspondences a plain-modern translator cannot buy at any price. ES-20260724-fluency-and-foreignness should note that this is its usual trade running backwards — here the foreignising choice is also the more accurate one at the level of sense.

And the cell partly relieves the conflation this page keeps complaining about. Revision trigger #3 asks for a structurally near, lexically non-overlapping pair, because both C3-bearing cells had proximity and overlap moving together. A-beowulf-ingeld separates them from the other side: Old English is structurally far from modern English (case, gender, strong/weak adjectives, no articles, OV/V2 order — the lead's log records articles and apposition as forced problems) while lexical overlap is partial. Register flattening by false friend occurs anyway. Overlap, not structural proximity, is the mechanism — which is what C3 claimed. Trigger #3 is not discharged (its own terms are unmet), but the alternative it was guarding against is now much less likely.

One qualification, and it costs C3 something. C3's headline is that overlap trades mediation cost for register cost. This cell has partial overlap, heavy register cost from false friends, and heavy mediation cost (ten forced forks). There is no trade here; both bills arrive. The Baudelaire trade happened because near/near meant the pair was culturally near as well as lexically overlapping. So: overlap adds register cost; it does not by itself subtract mediation cost. Stated as a correction, not a hedge.

Confidence: (a) rises to moderate-to-strong — three cells, three mechanisms (borrowing, synchronic drift, diachronic drift), seven translators, and now a stated mechanism with a within-root control rather than a pattern. The trade formulation is downgraded as above.

C4 — Handling one class of items several ways in one short text is normal, not a defect of one translator

Confidence: strong that the pattern recurs; open as to what it means.

A-shaw-spider-thread catalogued Shaw's four-way handling of one class of Buddhist toponyms and read it as a consistency cost — "a precedent worth not reproducing wholesale." Two more anchors, and the reading needs revisiting:

• Shaw: four handlings of one realia class in 1,000 words (substitute / transparent calque / copy-opaque / copy-opaque+incidental scaffold).
• Garnett: two dogs in a single sentence, two handlings ("Kashtanka" transliterated opaque; "Eel" translated for sense) — plus, across the story, the entire range including a measure conversion.
• Baudelaire: three italicised foreignisms, three outcomes (italics kept and voided; italics dropped; italics dropped).

Three translators, three pairs, three eras, three languages of origin, and every one of them is locally inconsistent within a coherent class. Either (a) consistency is a much weaker norm in published practice than the typology's definition implies — translators optimise item by item and let the class fall where it may — or (b) all three translations share a real defect. The project should not pick (b) from an armchair, and this is a question jury calibration could actually settle: a calibrated jury scoring consistency on class-inconsistent versus class-uniform handling of the same realia set would discriminate between the readings.

NARROWED 2026-07-25 (S017). A fourth translator and a fourth pair, and the pattern both survives and acquires a condition.

• The lead is inconsistent, inside one sentence: 『唐守』『藐姑射の刀自』『かくや姫の物語』 come out as Karamori (transliterated), The Old Woman of Hakoya (half calqued, half transliterated), The Tale of Princess Kaguya (calqued). Three handlings, one class, one sentence — the same shape as Garnett's two dogs.
• Both intralingual translators are perfectly uniform: all three titles copied.

So C4 holds only where a fork exists. Where the pair permits free copy there is nothing to be inconsistent about, so intralingual cases cannot corroborate C4 — they can only fail to contradict it. C4 is a claim about forked classes, not about classes as such.

CORROBORATED 2026-07-25 (S018), by an intralingual cell that does fork. A-beowulf-ingeld is the case S017 said could not exist — intralingual and forked — and C4 recurs in it at the sharpest resolution yet. One item, four translators, four handlings: lind-plega comes out as copy-opaque (Morris, "the lind-play"), calque retaining the wood (Gummere, "the linden-play", footnoted), omit-and-substitute (Kirtlan, "in the battle") and calque substituting the object (lead, "the shield-play"). bill gets one copy and three different substitutions ("brand", "the sword", "the blade"). And within single texts the class stays non-uniform: Gummere calques here-grīma and lind-plega but substitutes for wīg-bealu and bill; the lead calques four battle-realia and substitutes for the fifth. Seven translators, five pairs, five eras — none of them uniform on a forked class. Reading (a) — translators optimise item by item — keeps getting more expensive to deny; the calibrated-jury test named below is still the thing that would settle it.

A handling the catalogue does not have. Kirtlan renders eorlum on ende ("to the nobles in order / along the line") as "to the earls at the end of the high table." The high table is a later medieval English hall institution, absent from the source and from the society it describes. This is not substitution of a source item, not calque, not copy, not scaffold: it is domestication by supplying furniture — adding target-culture realia the source never had — and it is undetectable without the source open. Proposed to goodness-senses.md alongside measure conversion and preserved-inert.

And it gains the process evidence it was missing. T-genji-yomogiu-R04-v1's translator's log (§2c, frozen before any other rendering was read) records the lead noticing the inconsistency while committing it and declining to regularise, for a reason local to each title: the first two are lost tales whose titles mean nothing to anyone, so Englishing them would manufacture a false transparency; the third is a text an English reader may know, so transliterating it would hide what the source does not hide. That is direct evidence for reading (a) over reading (b) — the inconsistency was not inattention but three defensible local optima — from inside a translator rather than inferred from a finished text. It is one translator, self-reporting, so it does not settle C4; the calibrated-jury test named below remains the thing that would.

Two smaller by-products of the sweep, both proposed to goodness-senses.md:

• Measure conversion as a seventh handling (пуд → "forty-pound"): not substitution of a cultural item, not calque, not copy — a silent recalibration into the target's system, invisible by design and the most thoroughly domesticating move in Garnett's text.
• Preserved-inert as a named outcome rather than a handling: form retained, function lost. It appears twice, by different mechanisms — transliteration keeps the Иван/Ванька contrast as two strings and loses what the contrast means; italics keep Poe's foreignness-marker and lose what it marked. Worth naming because it is what "faithful retention" often actually delivers.

What this means for the framework

The operational consequence is one sentence: the goodness senses are not equally weighted across language pairs, and a framework that fixes their weights from J→E evidence will mis-specify every other pair. Concretely, on the three cells now in evidence:

• far/far → cultural-mediation heavy, marked-register loss heavy, syntax re-architected;
• far/near → register and expressive-morphology loss heavy, cultural-mediation cheap-with-a-copy-opaque-residue;
• near/near → cultural-mediation idle, style-correspondence heavy via cognate flattening, syntax nearly free, foreignisation possibly unavailable;
• zero/far (intralingual, added S017) → cultural-mediation elective, not forced — the load is whatever the translator chooses to take on, and two translators of the same language took on very different amounts; style-correspondence heavy via drifted rather than cognate forms, with lexical threads harder to hold than at zero overlap; grammar-borne meaning preserved or discarded at will, because the category is available; and the pair-indifferent, reader-distance problems fully intact.
• far/far with partial inheritability (diachronic, added S018) → cultural-mediation heavy and forced, exactly as at far/far with none; style-correspondence heavy from both mechanisms at once (marked-register loss and false friends in the drift window); consistency exposed, because every realia item is a fork; and the translator's declared archaism level partly determines which words are reachable at all.

The row above is why the profile now needs three columns rather than two. Rows one and five occupy the same two-axis cell and differ substantially, and what differs is inheritability. A framework that declared a pair profile from structural × cultural distance alone would give A-shaw-spider-thread and A-beowulf-ingeld the same profile and mis-specify the second on false friends — which is most of what a translator of that pair actually spends time on.

That fourth row is the one with the sharpest framework consequence, and it is a negative one: on this cell, the pair profile does not determine the sense weights, because the translator's purpose does. A framework that declares a pair profile and derives weights from it would mis-specify Yosano and Shibuya identically, and they differ from each other about as much as either differs from the English. Where inheritance is free, purpose-fit dominates the profile. This is directly relevant to D-20260724-04, whose ratified split already forbids deriving weights from the profile while this page is draft; it is now the first evidence that the split was the right call rather than a cautious one.

A framework release should therefore carry a pair profile — a declared reading of the pair on both axes and a statement of which senses that makes load-bearing — rather than a single fixed weighting. Because that is a methodological commitment with real downstream consequences, it was opened as a decision (D-20260724-04-pair-relative-sense-weights) rather than assumed here; a later session ratifies or rejects it (charter §8). No framework release exists yet, so nothing is retrofitted.

RESOLVED 2026-07-25 (S014), RATIFY-WITH-AMENDMENT. The decision was ratified as a split: framework releases must declare the pairs each recommendation is evidenced on and mark untested any cross-pair carryover (both standing requirements now), but are not required to weight recommendations from this two-axis profile — the profile is permitted as declared, provisional descriptive metadata only, while this page remains draft. Weighting becomes mandatory only after this page clears verification trigger #5 (a second-reader check of the three readings), a non-J→E workshop result exists, and the near/near confound is separated (or the empty cell filled). See the resolved decision page for the full record.

Second consequence, for the jury: the instrument's per-sense probes were designed against J→E problems. The within-class handling probe (substitute / calque / copy-opaque / scaffold, on the load-bearing gradient) transfers to all three cells and should be kept. What does not transfer is any assumption that cultural-mediation is where the action is.

What this does NOT establish

• Three translators, three short texts, ~~one reader~~ one reader plus one verification pass. Every cell is still n=1, and the readings are the lead agent's (internal-judgment-only). Updated 2026-07-25: the readings have now been second-read (RS-20260725-anchor-verification) — which is why C1 is a claim short. Second-reading is not replication: it checks that the reader read the text correctly, not that the pattern generalises.
• ~~C1 has no instance with English as the source (since the 2026-07-25 retraction), so it cannot yet distinguish "category mismatch in either direction" from "properties of translating into English."~~ DISCHARGED 2026-07-27 (S038) by A-sasaki-kuroneko (EN→JA, no lexical overlap, same source text as the retracted row). Its own limits, in its place: one passage, one published translator, one reader, not second-read; period confounded with translator, because the lead deliberately wrote a modern register against a prewar comparator; and the between-cat leg of the pronoun finding is 0 of 2 tokens.
• The externally anchored content is smaller than it was. Wallaert (2012) anchors Baudelaire's neologising pattern; the anchor's own instances of it (hyperdiabolique, antihumain) were withdrawn 2026-07-25 when a dictionary showed both are attested standard French.
• The axes are ordinal and arguable, not measured. "Russian is culturally nearer to English than Japanese is" is an argument from the realia these particular texts contain, not a metric. A different Chekhov story — the Orthodox-ritual-heavy ones — might place the same pair differently. The model would be much stronger with an operationalisation of the axes; it does not have one.
• The design is less incomplete and less Euro-weighted than it was, and is still both. (Updated 2026-07-25.) The structurally-near/culturally-far cell is filled, and with a non-European target. Remaining: the far/near-reversed cell (Chinese→Korean or similar); ~~English as a source whose target lacks an English category~~ — filled 2026-07-27 (S038), A-sasaki-kuroneko; and the fact that most pairs still have English at one end. A new gap opens in its place: every C1 instance is now a pair with zero or partial lexical inheritability, so nothing separates "category mismatch" from "category mismatch between distant languages". A structurally near, lexically non-overlapping pair (Finnish↔Estonian, German↔Dutch) is now C3's and C1's highest-value cell.
• Lexical overlap is conflated with structural proximity in the near/near cell — and also in the intralingual cell, differently: there both are extreme simultaneously, so A-yosano-yomogiu cannot separate them either. C3's named falsifier (a structurally near, lexically non-overlapping pair — Finnish↔Estonian, German↔Dutch) is still the fix, and is now the highest-value item for C3.
• Period is partly controlled, on one cell only. (Updated 2026-07-25.) Three of the four anchors are 1857–1939, all pre-professionalisation. A-yosano-yomogiu carries the project's first contemporary comparison text (Shibuya), which is what makes its period contrast possible — but only inside the intralingual cell, and there period is confounded with skopos: Shibuya's rendering is a scholarly crib published beside the classical text, Yosano's is a free-standing literary work. Almost every Yosano-vs-Shibuya contrast in that anchor is therefore not distance without being therefore purpose, and the anchor says so.
• ~~The intralingual cell rests on one passage of one work in one language. 951 characters, one chapter, one reader. Whether its findings are about intralingual translation or about Japanese is untested; a Beowulf or Analects cell would tell.~~ TESTED 2026-07-25 (S018). The Beowulf cell was built and the findings are not about Japanese: the inheritability formulation transferred and the intralinguality reading was refuted (C2). But the new cell has its own version of the same limitation — 56 lines, one poem, one reader — and the two intralingual cells are now two n=1 points, not a replication of each other.
• Period is uncontrolled inside the new cell, and this is its chief weakness. Morris 1895, Gummere 1909, Kirtlan 1913: three renderings inside nineteen years, all pre-professionalisation, plausibly all aware of each other. A-yosano-yomogiu had a contemporary comparator and this cell has none, because every modern Beowulf is in copyright. So the Morris-vs-the-rest contrast is clean, and the three-subject block as a whole is confounded with its period. Fixable with one copyrighted rendering read in brief excerpt and logged; not done.
• "Intralingual" is doing contested work. Whether Old English → modern English is intralingual translation, or interlingual translation between two languages sharing a name, is exactly what a linguist would press. The S018 result is arguably that the question is the wrong one — inheritability and reader distance predict the behaviour and the label does not — but the page has not earned the right to dismiss the objection.
• The third axis has four data points and no operationalisation. "Partial inheritability" is a judgment about a word-list, not a measurement, and it is the same weakness the first two axes have.
• One of the four subjects is the lead agent's own translation, at contamination: high. Its convergences with the two published renderings are evidence about translation problems only to the degree they are not evidence about the lead's training data, and that degree is unmeasured. The 煙絶えて convergence is the finding most exposed to this.
• No claim about quality. All three are precedents, not paragons; Garnett's reception record is largely hostile, and Baudelaire's celebrity is context, not evidence.

Revision triggers

Revise or retract if:

1. ~~A fourth precedent anchor in a new cell contradicts C2 — particularly a structurally near, culturally far pair whose realia load is heavy.~~ FIRED AND DISCHARGED, 2026-07-25 (S017). A-yosano-yomogiu is the fourth anchor, in the named cell, and it did contradict C2 — in the direction the trigger's wording did not anticipate. The trigger expected a heavy realia load; the load was light, and light for a reason that shows C2 was tracking a proxy. C2 is corrected in place (inheritability, not cultural distance; elective, not forced) and carries a new falsifier. The trigger is discharged for this round and remains standing against the corrected formulation. Note for future trigger-writing: this one nearly failed to fire because it specified a direction. A trigger should name the observation that would matter, not the outcome the author expects.
2. A contemporary translation shows systematic rendering of a target-lacking grammatical category — C1's "costs systematicity" clause fails. (Still standing, 2026-07-25. Shibuya's systematic keigo is not an instance: modern Japanese has the category, so it is the control C1 lacked, not the falsifier. See the C1 amendment.)
3. A near-pair anchor without heavy lexical overlap shows the same register flattening — C3's mechanism is misidentified. (Still standing on its own terms, but much less pressing after 2026-07-25 (S018). A-beowulf-ingeld separates overlap from structural proximity from the other side — structurally far, lexically partly overlapping — and register flattening by false friend occurred anyway, which is evidence that overlap is the mechanism. The trigger's own case, structurally near and lexically non-overlapping, is still unbuilt: Finnish↔Estonian, German↔Dutch.)

STANDING OBLIGATION, MOVED HERE 2026-07-28 (S049) FROM wiki/backlog.md, WHERE IT HAD REACHED THE REVIEW-OR-RETIRE AGE OF 10 WITH NO ARM TO SCHEDULE IT INTO. T4 has no live arm and the row is too load-bearing to retire, so it moves onto the page whose claims it would falsify — the S047 precedent for a row with nowhere to go, and the reason it is here is that a falsifier belongs with the claim rather than in a queue a rotation rule may never reach. The unbuilt cell is a structurally near, lexically non-overlapping pair — Finnish↔Estonian or German↔Dutch — and since 2026-07-27 (S038) it tests two claims rather than one: with A-sasaki-kuroneko filling C1's English-as-source cell, every C1 instance is now a zero- or partial-inheritability pair, so nothing in the anchor set separates "category mismatch" from "category mismatch between distant languages". One cell would separate C1's confound and test C3's mechanism at the same time, which makes it the highest-value single anchor this page could still gain. It travels with this page from now on and cannot be lost by a rotation rule. 6. NEW 2026-07-25 (S018). A translator working inside the drift window takes the reflex where the modern sense is impossible, or refuses it where the modern sense is merely slightly off — the window formulation of C3(a) is wrong, and reflex acceptance is tracking something other than local plausibility. The cheapest test is a second diachronic pair with reflexes at both ends of the drift range: classical→modern Chinese, or Middle→modern English. 7. NEW 2026-07-25 (S018). A heavily archaizing translation fails to reach source-internal threads a plain-modern one reaches, or a plain-modern one reaches them without coining — "archaism buys inheritability back" is misdescribed, and Morris's 9/10 is about Morris rather than about archaism. One further archaizing rendering of any diachronic pair would test it. 4. A calibrated jury rewards class-uniform handling over class-inconsistent handling — C4 resolves toward reading (b), and Shaw's original consistency-cost reading was right after all. 5. ~~Any of the three close readings is found to misread its source text.~~ FIRED AND DISCHARGED, 2026-07-25. The verification pass ran (E-20260725-anchor-verification → RS-20260725-anchor-verification): a deterministic audit of all 209 quoted correspondences across all three anchors (207 attest, 0.990), a blind arm, an adjudication arm with planted decoys, and external dictionary checks. The trigger fired: one reading was found to misread its source text — the Poe him/it claim — and C1 lost an instance and its confidence rating (above). Two further claims were softened and one uncatalogued departure was added to the Shaw anchor; the Garnett reading came through intact at 85/85. The trigger is discharged for this round and remains standing: any further misreading found re-fires it.

Change log

• 2026-07-28 (S044) — C1 gains a scope condition from a non-Anglophone primary, and it is the first bound this page has taken from outside the Anglophone base (S-luxun-yingyi, Lu Xun's 硬譯 essays 1930–31, read whole in Chinese; paired with T-yingyi-jiejixing-R04-v1). The change is one clause and it is not a retraction: C1 presupposes the target's inventory is fixed, and Lu Xun's programme is the deliberate denial of that presupposition. Where a translator treats the target's grammar as mutable and is importing syntax on purpose, the "cost in systematicity" is the intended mechanism, and C1 describes the wrong thing. No claim of C1's is withdrawn and no confidence rating moves; what changes is that a framework recommendation derived from C1 now has to say which kind of translator it is addressed to. Recorded in place at the end of the C1 section, with the corpus test that would settle Lu Xun's own prediction named as out of the project's reach. Page stays draft, internal-judgment-only.
• 2026-07-27 (S038) — C1's named gap is filled and C1 survives it; the loss clause returns to moderate-to-strong, and a bias under the log corpus is exposed (A-sasaki-kuroneko, EN→JA; paired with T-black-cat-R04-v1, a lead translation of the same seven paragraphs whose 24-point log was frozen at 505520a before the Japanese comparator was fetched). Changes: (1) C1's English-as-source instance exists, on the same source text as the row retracted in S015, in a pair with no lexical inheritability — 佐々木 repays English's third-person animacy at 5 of 15 cat-referring masculine sites with 彼, and places all five where the cat is an agent, tracking Poe's own shift to it at the mutilation and the hanging. Intermittency is the claim and intermittency is what was measured. (2) Confidence on the loss restored to moderate-to-strong; the conjecture that "the regularity is about category mismatch, not about which language is strange" is now reached by evidence and holds. Revision trigger #2 does not fire — one site in three is the opposite of "deployed at every occurrence". (3) Forced explicitation is asymmetric, and the asymmetry looks structural: a category the target obligatorily has forces a decision at every occurrence; a category the source has and the target lacks presents no slot, so the information goes silently. Two sites in 880 words out of English, against "every clause" and "every count noun" into it. (4) Therefore RS-20260727-log-typology's C4b rate is partly a fact about English being the target — twenty of twenty-one logs translate into English. Recorded as a limitation on that result. (5) C2's elective/forced distinction recurs for a grammatical category: at It was **a** black cat 佐々木 takes Japanese's free numeral-classifier option and the lead refuses it, with the log giving the reason; identical demonstrative budgets, opposite distributions. (6) Two negative results for C3 read the other way. Poe's Latinate-over-Germanic register transferred cheaply at zero inheritability (Japanese 漢語/和語 is a two-storey lexicon of unrelated origin), where the maximal-overlap French pair flattened it — two-storey-ness is not inheritability. And on the passage's one repeated word (indefinite ×2) 佐々木 holds the thread 2/2 and the lead holds it 0/2, both at zero overlap, which costs the project the reading that the lead's 3/3 on うるはし was because it had no reflex to refuse. (7) Anchor table grows a sixth row. Page stays draft, internal-judgment-only.
• 2026-07-25 (S018) — a second intralingual cell separates inheritability from intralinguality; C2's correction survives its first independent test, C3(a) gains a mechanism, and the model grows a third axis (A-beowulf-ingeld; paired with T-beowulf-ingeld-R04-v1, a lead translation of the same passage whose freeze was committed to git before any published rendering was fetched). Changes: (1) C2 tested and survived — Old English → modern English is intralingual but not inheritance-rich, and behaved like an interlingual pair: ten forced realia forks in 56 lines, all four translators forking on all ten. The rival reading of the S017 result — that intralinguality as such makes mediation elective — is refuted. The free-copy residue is substances, drinks and kinship; institutions and material culture fork. (2) C2 repaired on opacity — S017's "intralingual copy never manufactures opacity" is too strong; Old English names are transparent common nouns (Wiðergyld = requital) and all four translators copy them into opaque strings. The variable is reader distance, not the language boundary, which unifies C2 with the pair-indifference observation at C3. (3) C3(a) sharpened into a mechanism with a within-root control — reflex acceptance is 0/12 at total drift and 19/40 at partial drift; glæd (2026) is taken by 3 of 4 translators and gladiað (2037), same root eleven lines later, by none, because the modern sense is wrong at the first site and impossible at the second. The false friend operates in a window. This retro-explains why the zero-overlap translator held the うるはし thread 3/3 in the Japanese cell: no reflex and a totally drifted reflex are the same condition. (4) A second C3 mechanism named — archaism buys inheritability back: Morris takes 9/10 partial-drift reflexes against 4, 5 and 1, and the only translator to hold the source's duguð/dugan etymological thread complete, 3/3. Foreignisation on a diachronic pair is a capability, not only a stance. (5) C3's trade formulation downgraded — this cell carries heavy register cost and heavy mediation cost; overlap adds register cost without subtracting mediation cost. (6) C1 extended — forced explicitation arises from an obligatory target category as readily as from a lost source one (wīf → woman/wife; articles). (7) C4 corroborated in the case S017 said could not exist — an intralingual cell that forks: one item, four handlings; and added realia ("the high table") proposed as a handling the catalogue lacks. (8) The two-axis model is now three-axis; the far/far cell contains two anchors that differ mainly in inheritability. Two new revision triggers (#6, #7). Page stays draft, internal-judgment-only.
• 2026-07-25 (S017) — the empty cell is filled, by intralingual translation, and all four claims move (A-yosano-yomogiu; paired with T-genji-yomogiu-R04-v1, a lead translation of the same passage whose log was frozen before either published rendering was read). Revision trigger #1 fired and is discharged. Changes: (1) C2 corrected — mediation load tracks inheritability of the source's signifiers, not cultural distance alone; where inheritance is free the fork is elective, and two translators of one language elected oppositely (Yosano 5 substitutions, Shibuya 1). Cultural distance predicted the load in the three interlingual cells because, between languages, distance and non-inheritability move together; intralingual translation pulls them apart. New falsifier attached. (2) C3 extended and upgraded to moderate for (a) — the false friend need not be a cognate; semantic drift over time does the same work, demonstrated on うるはし/麗しい and めざまし/目覚ましい. The うるはし thread was held 3/3 by the translator at zero overlap and 0/3 by the translator at maximum overlap, because overlap supplies a tempting candidate that must be refused site by site. (b) narrowed to evacuated markers; intralingual translation widens the foreignisation choice rather than foreclosing it. (3) C1 gains a control and a second half — the same category rendered into a target that has it (Shibuya, systematically) and one that does not, plus a translator who had it and dropped it; and forced explicitation named as the second cost, the lost category making the translation more determinate than the source. (4) C4 narrowed to forked classes, and given its first process evidence: the lead's log records noticing the inconsistency while committing it, for three defensible local reasons. (5) Framework consequence added: on this cell the pair profile does not determine sense weights because purpose-fit dominates, which is the first evidence that D-20260724-04's ratified split was right rather than merely cautious. Page stays draft, internal-judgment-only.
• 2026-07-25 (S015) — second-reader verification of the three underlying close readings; C1 loses an instance and a confidence grade (RS-20260725-anchor-verification). Revision trigger #5 fired and is discharged. Changes: (1) C1's Baudelaire row retracted — Poe does not split the two cats by pronoun; C1 drops from three instances to two, from "strong" to "moderate" on the loss, and now has no instance with English as the losing source, so the paragraph claiming the regularity is "about category mismatch, not about which language is strange" is downgraded to a conjecture the evidence does not reach, and filling that gap is named the highest-value single addition to C1. (2) The Baudelaire coinage instances behind C3's repair mechanism are withdrawn on external-dictionary evidence (the Tier-2 Wallaert pattern is untouched). (3) C1's Shaw row is unaffected but the mimetics sub-claim behind it was softened at the anchor. (4) C4 unaffected; C2 unaffected. Nothing here upgrades the page: it stays draft, internal-judgment-only. This satisfies revisit condition (a) of D-20260724-04 — the other two conditions (a non-J→E workshop result; separating the near/near confound) remain open, so pair-profile weighting stays optional descriptive metadata.
• 2026-07-24 (S013) — created, draft. First theory page. Synthesises the project's three precedent anchors into a two-axis model (structural distance × cultural-referential distance) with four claims: C1 grammaticalised categories are lost and repaid lexically at a cost in systematicity (3 instances, including one with English as the losing source); C2 cultural-mediation load tracks the cultural axis, not the structural one; C3 lexical overlap trades mediation cost for register cost and can foreclose the foreignisation choice; C4 within-class handling inconsistency is normal across all three anchors, which reopens the consistency reading S012 gave it. Proposes measure conversion (seventh handling) and preserved-inert (an outcome) to goodness-senses.md; supplies a prior question to ES-20260724-fluency-and-foreignness; opens D-20260724-04-pair-relative-sense-weights.