Repository path: workshop/experiments/E-20260728e-venuti-specifiability/design.md · rendered 2026-09-09

E-20260728d — can a foreignizing strategy be specified in advance?

Frozen 2026-07-28 (S048), before the source was read for translation and before either translation existed. The commit of this file and of R07/R08 is the freeze.

Question

Lawrence Venuti's The Translator's Invisibility (Routledge, 1995) supplies the project's vocabulary for the domestication ↔ foreignization axis. Maria Tymoczko's standing objection is that he supplies no operational criteria for what counts as foreignizing. Venuti's own text says something adjacent and stronger about his own method: the relationship a resistant translation establishes with its source "cannot be calculated before the translation process is begun," and its techniques "more often than not […] surface accidentally as possibilities are tested, their effects evaluated only after the fact, when rationalization occurs" (ch. 6, p. 296).

Question. Write both poles down as rule sets, in advance, from Venuti's own descriptions; translate one passage under each; and count what proportion of contested decisions the rules actually decide.

This matters to the project and not only to Venuti. S-venuti-invisibility has since S006 recommended importing the pair as "a scored axis / pair of tendencies." An axis that cannot be specified cannot be scored, and the jury instrument would be scoring something else under its name.

Materials

• Source. I. U. Tarchetti, «Un osso di morto», from Racconti fantastici (Milan, 1869) — public domain, Italian Wikisource, transcribed from the 1869 djvu. Opening two paragraphs, 362 words, stored at workshop/translations/osso-di-morto/source-it.txt.
• Why this text. Tarchetti is the writer Venuti's ch. 4 builds its dissidence argument on and whose Fantastic Tales Venuti himself translated (Mercury House, 1992). Italian is Venuti's own working pair. The prose is periodic, dash-parenthetical, and archaic-literary — the kind of source on which the two poles should visibly diverge if they diverge at all. Italian is not new to the project — ARM-longwork completed Verga's «Jeli il pastore» at S047 — but this is its first Italian passage rendered twice under contrasting regimes.
• Regimes. R07 (fluency, ten rules F1–F10) and R08 (resistancy, ten rules R1–R10), both frozen before translating.

Contamination — declared before selection, per the standing rule

suspected, and unmeasurable. Venuti published his own English of this tale in Fantastic Tales (Mercury House, 1992); it is in copyright and not freely reachable, so no longest-common-run measurement against a comparator is possible. This is the sixth lead translation to carry an unmeasured declaration and the standing rule names that as a placeholder failure mode; it is accepted here, with the cost stated, because no arm of this design compares the renderings to any published English. The measurement is over the project's own rule sets and its own decision log. A later session that obtains Venuti's rendering can run the measurement retrospectively; nothing here would have to be redone.

Priming: the lead read The Translator's Invisibility whole in chs. 1, 6 and 7 and in part elsewhere before translating — deliberately, since the rule sets are drawn from it. The book contains no English rendering of this tale, verified by string search over the full extracted text (osso di morto, Dead Man, Letter U: 0 hits each). Venuti's English of Tarchetti's essay «Idee minime sul romanzo» was read (ch. 4 pp. 152–161); that is a different text and is not the source here.

Procedure

1. Freeze R07, R08 and this design. Commit. (The freeze.)
2. Translate the passage under R07. Log every contested decision with its coverage code D/P/S and rule numbers.
3. Translate the passage under R08, from the source — not from the R07 output. Log as above, plus the INT flag where a decision was made on an interpretation of the source's meaning rather than by a rule.
4. Freeze both translations and both logs. Commit.
5. Compute the outcome measures below.
6. Independent check. One call to a non-Anthropic panel model (config/models.md), given the two rule sets, the source, both renderings and the bare decision sites — not the lead's codes — and asked to assign D/P/S independently. Report agreement.

Outcome measures

• cov(R) — coverage. Of the contested decisions logged under regime R, the proportion coded D.
• sep — separation. Of the decision sites that appear in both logs, the proportion at which the two renderings differ in wording.
• int(R08) — the count of R08 decisions flagged INT.
• Length. Words in each rendering.
• agree — proportion of sites where the independent classifier's code matches the lead's.

Predictions (registered; all five may fail)

• P1. cov(R07) > cov(R08). Rationale: Venuti states the fluency features as a checklist derived from a review corpus; he states resistancy as something that surfaces accidentally and is rationalized after the fact.
• P2. cov(R08) < 0.50 — fewer than half of the contested resistant decisions are settled by rules written in advance.
• P3. sep > 0.50 — the two regimes are separable in output, not only in rationale. A failure here would mean the rule sets are ornamental.
• P4. int(R08) ≥ (number of R08 sites coded D) — interpretation does at least as much of the deciding as the rules do.
• P5. The R07 rendering is longer in words than the R08 rendering. Rationale: F7 supplies subjects, verbs and connectives that R5 forbids. If false, the length confound the project measured in S046 runs the opposite way for this pair and the fluency/foreignization contrast is not a length contrast.

Failure criteria

• If fewer than 15 contested decisions are logged under either regime, the coverage rates are not reportable and the run is recorded as underpowered.
• If sep ≤ 0.25, P3 fails hard and the whole design is reported as having failed to produce two conditions; the coverage numbers are then descriptive of one translation written twice and must not be cited as being about foreignization.
• If the independent classifier's agree is below 0.60, the lead's coverage rates are reported as unreliable and the headline numbers become the classifier's, not the lead's.

What this cannot establish

• n = 1. One passage, one language pair, one translator, one pair of rule sets assembled by the party being tested. Nothing here generalises; it is a case demonstration with a number attached.
• The target-culture half of Venuti's theory is not in the design at all (R08 §Known limitations). Foreignization for Venuti is contingent on a target-language situation; two rule sets operating on a source-to-target mapping cannot see that, so a low cov(R08) is evidence about the discursive half only.
• No claim about translation quality is made or licensed. The lead never judges its own translation (charter §5); no quality evaluation is part of this design.
• The pre-run critic pass was not run, in departure from continue-prompt.md §6. The reason is stated rather than smoothed: the design's known weakness is self-scored coverage, and the session's one budgeted call was spent on an independent post-hoc classifier (step 6) instead, which attacks that weakness directly. A critic without access to the book could not have checked the only other soft joint, the rule assembly.