Repository path: workshop/experiments/E-20260802b-pole-uniformity/design.md · rendered 2026-09-09

E-20260802b — is a translator's position on the domestication↔foreignization axis a stance or a running trade?

Frozen 2026-08-02 (S088) before the locus was translated and before any published English of the locus was read. ARM-fluency-record step 1, study limb.

1. Why this question, and why it is not a question about the apparatus

wiki/findings/sense-dossier.md records the sharpest unanswered Tier 2 challenge in the typology: the wording of naturalness is Nida's, and The Translator's Invisibility ch. 1 is an argument against it. Logged S006, sharpened S048, never answered — and E-20260801g found naturalness on one side of 15 of 16 purpose-borne trades, so a sense doing that much work cannot keep an unanswered challenge indefinitely. ARM-typology-derivation step 3 owes the motion and cannot open it, because the motion needs domestication↔foreignization evidence the project has never assembled.

The motion's shape depends on a prior fact nobody has measured. Venuti's argument, and Schleiermacher's before it, treat the axis as a stance — a single strategic position a translator adopts and then executes. Schleiermacher is explicit that the two methods are not a spectrum: one "must be pursued as strictly as possible", and mixing yields "a thoroughly unreliable result" (S-schleiermacher-methoden §C). Venuti's domesticating and foreignizing name whole translations, not choices. The project's own strongest cross-corpus finding says the opposite about every other axis it has looked at: seven translators, five language pairs, five eras, and not one of them uniform on a forked class (sense-dossier.md, consistency C4).

These make opposite predictions about the same published texts, and the texts are free.

The subject-rule sentence (wiki/tracks.md, continue-prompt.md §4.5): this unit teaches whether "foreignizing" and "domesticating" name something a literary translator does once, for a whole work, or something decided again at every hard word — which is the difference between naturalness being a norm one accepts or rejects and naturalness being a trade one makes site by site. That is a claim about translating literature, not about this project's instruments.

2. Materials — all public domain, original and translations both free

Source. Iliad XII.310–328, Sarpedon to Glaucus. Greek from the Monro–Allen OCT as carried by the Perseus canonical text (tlg0012.tlg001.perseus-grc2), stored at materials/source-grc.json. Homer PD; Monro (d. 1905) and Allen (d. 1950) — the 1920 edition is PD in the US on the pre-1929 publication basis.

Why this passage. It is the passage the Arnold–Newman controversy of 1861–62 argues over twice — the one place where the English-language debate about foreignizing and fluency, conducted first-person by both parties, has a shared text under it. It is a complete speech, 19 lines, and every major English Iliad renders it.

The six published renderings, spanning 287 years and both poles of the controversy:

The two criterion statements, quoted from the free PD volume On Translating Homer (Arnold and Newman, Routledge; Project Gutenberg #65381), logged in wiki/base/consulted.md:

• Newman, stated by himself and quoted by Arnold: he "aims at precisely the opposite: to retain every peculiarity of the original, so far as he is able, with the greater care the more foreign it may happen to be"; the translator's "first duty is a historical one, to be faithful."
• The fluency position Newman states in order to condemn it: the translation should be such "that the reader should, if possible, forget that it is a translation at all, and be lulled into the illusion that he is reading an original work — something original from an English hand."

That first quotation is a first-person, universally-quantified claim by a translator about his own practice (every peculiarity, the more foreign the greater the care), made about the very book this experiment measures. It is the reason Newman is the registered positive control.

3. The lead's translation limb, and what it is for

Two renderings of XII.310–327 by the lead, from the Greek alone, under the two frozen regimes: T-iliad-sarpedon-R07-v1 (fluency) and T-iliad-sarpedon-R08-v1 (resistancy). They are subjects, not evidence about the published six: they exist to answer the complementary question — whether a translator instructed to hold a pole globally can do it — and to generate the site list from the inside, by a translator who had to decide each one, rather than by inspection of the published texts.

The locus is XII.310–327 for the lead and XII.310–328 for the published census, and the reason is recorded rather than hidden. While extracting the gate span (XII.331–341) from Butler and Lang, the extraction window opened on the closing line of Sarpedon's speech, so the lead saw line 328 in two published English renderings before translating. Line 328 is therefore excluded from the lead's renderings and from every comparison involving them; it stays in the published census, where no lead output is involved.

Contamination gate, run before the locus was translated (standing rule, CLAUDE.md): the lead rendered XII.331–341 from the Greek alone and it was compared against Butler and Lang by tools/dependence_check.py.

Zero shared 7-grams against either comparator, and the lead's longest runs (6, 5) are of the same order as the two published translators' own (4). Both figures are reported because run length alone has failed as a proxy three times (CLAUDE.md). contamination: none on the measured basis, on the most-translated text in existence — which is a datum for the standing finding that the comparator's fame is not what governs overlap.

4. The S073 binding condition on R07, discharged here before the source was opened

workshop/regimes/R07-fluency.md requires any design translating under R07 to settle two interpretive rulings and to name the missing rules in its own text, before it opens the source. Settled here:

• IR1 — F10's scope: taken LITERALLY. "Nothing that calls attention to the language" is read as requiring the plain rendering wherever one is available, including where the plainness costs a figure the source contains. This is the opposite of the ruling the three earlier runs took, and it is taken because RS-20260730e §2 measured the question: put to three non-lead readers at four figurative sites, 8 of 12 cells returned that F10 requires the plain rendering and none returned that it favours carrying the figure over. The project should follow its own measurement rather than its own precedent.
• IR2 — F6's scope: taken TEXTUALLY. F6 licenses departure from source word order and forbids nothing; paraphrase is not excluded by it. RS-20260730h §6 found the contrary reading unsupported by F6's text in all three earlier runs.
• M1–M6 are expected to fire and are logged, not repaired. M1 typography/layout, M2 sequence of tenses, M3 F4's heading being wider than F4's clauses, M4 holding a lexical choice steady across a text, M5 second-person address systems, M6 proper-noun knowledge as opposed to proper-noun convention. M6 is certain to fire here (Xanthos, Lycia, the temenos).
• The exclusion-test coverage rate is not computed. Only the strict §5 test (a rule names the feature and exactly one live option survives) is reported, with the mean number of live options per site beside it.

5. Procedure

1. Freeze this design (this commit), then translate the locus twice under R07 and R08, logs written as the drafts are written, frozen by commit before anything else happens.
2. Extract the six published renderings of XII.310–328 and file them on the anchor page A-homer-sarpedon.
3. Build the site list from the union of the lead's two logs — a site enters if the lead had two or more live renderings at it — and then keep only sites at which at least five of the six published translators are codeable.
4. Code every (site × translator) cell F / D / —: - F — the rendering keeps the source's form where English has a norm of its own: the compound stands as a compound, the litotes as a litotes, the Greek institution or name-form unassimilated, the source's clause order or discontinuity reproduced. - D — the rendering replaces the source's form with the target-language norm: the compound resolved, the litotes turned positive, the institution given an English equivalent, the name given its English convention, the period rebuilt. - — — the translator does not render the feature at all, or the cell cannot be decided.
5. Independent coding. The identical site list, the identical six renderings, and the F/D definition above and nothing else — no hypothesis, no prediction, no lead coding — go to two non-Anthropic seats in stateless calls (config/models.md). This is the S087 amendment A3 procedure: the critic pass showed a seat given the design in hand recovers the lead's labels at 15 of 24 and a seat given items alone at 9 of 24, so the seats are never shown this page.
6. Score by analysis/score.py; verify by analysis/verify.py, recomputing every reported number from the stored cells with mutation tests.

6. The quantities

For translator t over the n codeable sites, with f(t) = count of F and d(t) = count of D:

• Pole rate p(t) = f(t) / (f(t)+d(t)).
• Uniformity U(t) = max(f, d) / (f + d) — 1.0 is a pole held perfectly; 0.5 is a coin.
• Site consensus c(i) = max(F, D)/(F+D) across translators at site i.
• Variance split — Var_between-translator(p) against Var_between-site(c'…), both computed on the same cell matrix as the mean squared deviation of the row means and of the column means from the grand mean.

7. Registered predictions — the two theories make opposite calls

P1 is the primary. It is the one prediction with a named first-person claim behind it, made by the translator being measured, about the book being measured.

The lead's own limb, registered separately. U(R07-lead) and U(R08-lead) are reported but are not evidence for or against P1–P4: a translator executing a written rule set is not the same object as a translator executing a programme. They answer only can a pole be held when holding it is the instruction — registered prediction: both ≥ 0.90, since the rules are explicit and the translator is trying.

8. Failure criteria — registered, and binding

• FC1 — positive control. The coding must recover the one ranking every party to the 1861 controversy agrees on: p(Newman) is the highest of the six and p(Pope) the lowest. If it does not, the coding is not measuring the axis, and P1–P4 are withheld and only the census is reported.
• FC2 — power. At least 10 sites must survive step 5.3's five-of-six rule. Below that, the census is reported and P1–P4 are reported as underpowered.
• FC3 — the S087 lesson, and it gates the conclusion rather than the labels. The primary conclusion must reproduce on each independent seat's own coding, separately. If P1's verdict (U(Newman) ≥ 0.90 or not) differs between the lead's coding and either seat's, the primary is withheld and all three codings are reported side by side. Cell agreement is reported for context with its chance rate, and does not gate on its own — a binary code's agreement is ~50% by chance and a threshold on it would be theatre.
• FC4 — no post-hoc site list. The site list is fixed at step 5.3, from the lead's logs, before any published rendering of the locus is read. Sites may be dropped by the five-of-six rule and never added after the published renderings are open.

9. What a result licenses, and what it does not

• It does not license any change to wiki/goodness-senses.md. It supplies evidence for a motion a later session opens and a session after that ratifies (charter §8).
• It says nothing about which pole is better. Neither regime nor translator is scored for quality here, and no jury is involved. provisional: true, Tier D having answered NO at S086.
• It is six translators on one passage of one work. Whatever it finds is a finding about this passage until it is replicated on another — the standing weakness of every anchor this project has built, stated rather than discovered.
• The lead did the coding of record. RS-20260802-voice-warrant §6 measured what that is worth on a seven-way sense assignment (9 of 24). FC3 is the response, and it is the reason the conclusion rather than the labels is what has to survive.

10. Budget

Two independent coding calls, non-Anthropic seats, and one pre-run critic call. Worst case built from max_tokens (note (abc), and note (bhf) from S087 — a cap bounds reasoning as well as visible output): critic at 24,000 → $0.20; two coders at 16,000 each → $0.30. Declared worst case $0.50 against $3.77 of headroom on 2026-08-02.