Repository path: workshop/regimes/R07-fluency.md · rendered 2026-09-09

R07 — fluency (Venuti's domesticating checklist, executed)

v1.0 (frozen 2026-07-28, S048.) An R06 (lead, source-only, single pass) with one addition: a numbered rule set, written and frozen before the source is translated, which the translator must consult at every choice and against which every choice is scored.

The rule set is not the project's invention. It is Venuti's own description of what makes an English translation fluent, assembled from the review corpus in The Translator's Invisibility ch. 1 §I (pp. 4–5) and from his account of the formal techniques that produce transparency in ch. 6 §II (pp. 286–287). Venuti states these as the features a reviewer rewards; R07 turns them round and states them as instructions a translator can follow. Whether that turn is legitimate is exactly what the paired R07/R08 comparison measures.

Why it exists

The project imported domestication ↔ foreignization as a scored axis in S006 and could not say how anything would be scored on it. Maria Tymoczko's standing objection is that Venuti supplies no operational criteria. This regime and R08 test that objection by trying to write the criteria down and then execute them — separately for each pole, because the two poles are not documented to the same depth in the source, and R07 exists partly as the control: if the fluency rules also fail to decide, the finding is about rule-writing in general and not about foreignization.

Roles

• Translator — the lead agent, as a labeled subject. No panel call, no API cost.
• Judge — never the lead (charter §5).

Human involvement

None.

The rule set (frozen; cited by number in the log)

F1 — Current, not archaic. Where a word or construction has a current and a dated form, take the current one. F2 — Standard, not specialized. Prefer widely used vocabulary to technical or learned vocabulary, except where the source term is itself technical. F3 — Standard, not colloquial. No slang; no dialect. F4 — No foreign matter. No untranslated source-language words, no calques that read as foreign, no source-language proper-noun conventions left unadapted where an English convention exists. F5 — One dialect throughout. Do not mix British and American usage. (The regime does not choose which; the translator declares the one it is writing in and holds it.) F6 — Idiomatic syntax before close syntax. Where following the source's word order or clause order produces English that would be called "not quite idiomatic," depart from it. F7 — Continuous, easy syntax. No sentence fragments. Every clause has its subject and its finite verb. Connectives are supplied where English wants them and the source leaves them implicit. F8 — Fixed, precise meaning. Where the source is ambiguous or polysemous, choose one reading and write it. Do not reproduce the ambiguity. F9 — Rhythmic definition and closure. Sentences end; long periods may be broken. Avoid a trailing or unresolved cadence. F10 — Nothing that calls attention to the language. No word choice whose effect is to be noticed as a word choice.

The ten rules are Venuti's list, renumbered. Rules F1–F5 are the lexical half (ch. 1 pp. 4–5); F6–F9 the syntactic half (ch. 1 p. 5, ch. 6 pp. 286–287); F10 is the criterion Venuti quotes from Charles Bernstein — "anything that might concentrate attention on the language itself" — as the thing plain style excludes (ch. 1 p. 6).

Procedure

1. R06 §1 unchanged. Source only. The translator reads no published rendering of the same passage into any target language before the translation and its log are frozen. Priming, if any, is declared.
2. The rule set above is frozen and committed before the source is read for translation. The commit is the freeze. A rule may not be added, deleted or reworded once translating has begun; a rule discovered to be missing is recorded in the log as missing, and the version is bumped only in a later run.
3. Translate once, straight through, as R06 §2. No return pass.
4. The log is written as the draft is written, decision by decision, as R06 §3.
5. Every logged decision carries a coverage code, assigned at the moment the decision is made: - D (decided) — a numbered rule names the feature at issue and only one of the live options satisfies it. The rule number is recorded. - P (permitted) — one or more rules bear, and two or more live options satisfy all of them. The rules do not choose. Rule numbers recorded. - S (silent) — no rule bears on the feature at issue at all. A decision is logged when the translator was aware of two or more live renderings at the moment of writing. Sites where nothing was live are not decisions and are not logged; the log therefore under-counts total choices by construction and the coverage rates are rates over contested sites only.
6. Freeze before anything else happens (R06 §4). Commit the translation and log before any comparison, scoring or evaluation is designed.
7. Contamination declared with a one-line basis.
8. File under workshop/translations/<work>/R07-v<N>/.

Outputs

translation.md with its translator's log and the per-decision coverage codes. No run/. Cost $0; never ledgered (charter §6, A4).

Known limitations (by design)

• The translator scores its own rule coverage. The D/P/S code is assigned by the same agent that wrote the rules and made the choices. The code is defined to be a textual test — does a rule name this feature — rather than a judgment of fit, which is the most that can be done without a second reader; a second reader is a separate, budgeted check and is not part of the regime.
• Everything R06 lists, including that a real error survives because there is no return pass.
• The rule set is an assembly, not a quotation. Venuti nowhere prints ten numbered rules. Turning his description of a reviewers' norm into a translator's instruction is an interpretive act by this project, and if the assembly is wrong the coverage measurement is measuring the assembly.
• F5 does not choose a dialect and so cannot be violated by a translation that declares one and holds it. It can only be violated by inconsistency.

Changelog

• 2026-07-28 v1.0 (frozen, S048) — created, with R08, for RS-20260728e-venuti-specifiability. Frozen on creation for the same reason R06 was: the first comparison it takes part in is the comparison it exists to make.

Measured defects — added S064. The rule set above is UNCHANGED and still frozen v1.0

RS-20260730e-rule-coverage / E-20260730e. These are findings about how this regime's statistic has been applied, not edits to F1–F10. Nothing here bumps the version; a v1.1 is a question for a later run.

1. cov(R07) has been computed with the wrong test, and every published coverage rate is of unknown correctness. §5 defines D as a rule names the feature and only one of the live options satisfies it. In practice a D was recorded whenever a rule excluded an option, which is a different and much weaker condition. Demonstrated on three sites the lead had chosen as its clearest F4 cases: at two of the three, F4 excludes one option and leaves two English renderings standing, so the correct code by this page's own text is P. Three independent readers returned P there and the lead had returned D**.

Binding condition on use, from S064. Any page reporting a cov(R07) figure must either (a) state that its D codes were assigned by the exclusion test and are an upper bound, or (b) re-derive them against §5's exactly one survivor test and say so. ~~The three existing figures — 13.6% Italian, 36.1% French, 66.7% Bengali — all fall under (a) and none has been re-derived.~~ DISCHARGED S067: all three have now been re-derived under (b) — see §Measured defects 5 below. The exclusion-test figures may still be quoted, and only as upper bounds, and only beside the strict figure.

2. F10's scope is not settled by F10, and the ruling the two later runs were executed under has no independent support. IR1 (F10 governs the translator's latitude, not the reproduction of source figures) was put to three non-lead readers with F10's text verbatim at four figurative sites: 8 of 12 cells returned that F10 REQUIRES the plain rendering, 4 of 12 that the text does not settle it, none that it favours carrying the figure over. The registered refutation criterion (unanimity at ≥ 3 of 4) did not fire, so IR1 is unsupported rather than refuted — but the majority reading is the literal one IR1 rejected, and a v1.1 must say which reading it means.

3. Two missing rules recur across every run and two more are new. M1 typography/layout and M2 sequence of tenses were recorded on the Bengali run and recur unchanged on the French one — different language, script, century and genre — so they are properties of the list, not of a story. M3: F4's heading ("No foreign matter") reaches third-language matter and its clauses, which name source-language words, do not. M4: nothing in F1–F10 bears on holding a lexical choice steady across a text.

4. §Known limitations' first bullet is now measured rather than declared. "The translator scores its own rule coverage" — an attempt to have non-lead readers score it instead failed its own pre-registered positive control and its own byte-identical repeat control (self-agreement 0.757 against a 0.80 floor). So the limitation stands and is now known to be hard to remove, not merely unaddressed.

Re-derivation and four further defects — added S067. The rule set above is STILL UNCHANGED and still frozen v1.0

RS-20260730h-strict-coverage / E-20260730h, verifier 242 checks / 0 failures / 4 mutation tests all caught. Again: findings about how this regime's statistic has been applied and about what its ten rules do not reach. Nothing here bumps the version.

5. The three published coverage rates, re-derived against §5's own text. The binding condition above is discharged. All 39 published D codes were re-coded from the logs' own option sets and cited rule numbers, by a frozen procedure, and independently checked by a three-seat non-lead panel that passed both of the controls the S064 instrument failed.

The 4.9-fold spread this regime's statistic showed across runs is an artifact of the wrong test. Under §5's own text the spread is 5.8 points and the order inverts — the run with the lowest published rate has the highest strict rate, because its D codes are register decisions (F1, F2) where one option genuinely stands, and the Bengali run's are realia decisions (F4) where two or three acceptable English renderings always remain.

And the exclusion-test rate is rank-identical to the mean number of live options per site across all four runs (2.591 → 13.6%, 2.833 → 36.1%, 3.100 → 66.7%, 3.309 → 81.8%). cov(R07) computed by the exclusion test measures the length of the translator's option list. The strict figure does not.

6. TWO interpretive rulings, not one, and both were carried into the runs unsupported. IR1 (F10's scope) was already recorded. IR2 is new: all three published runs read F6 as licensing departure from word order but forbidding paraphrase, and excluded options on that ground. F6's text licenses departure and forbids nothing. IR2 is load-bearing — it turns D codes into P codes where it is applied. A v1.1 must say which reading of F6 it means.

7. Two more missing rules, and the count is now five. M5 — nothing in F1–F10 bears on second-person address systems; the vous → tu site was coded D on F4, and F4's clauses name untranslated words, foreign-reading calques and proper-noun conventions, none of which is a T/V system. M3 is confirmed as structural rather than incidental: on a Russian source carrying Ukrainian it fired three times in 955 words, and at each of the three F4's heading and F4's clauses give opposite answers.

8. An F5 D code is not the same kind of object as any other D code. F5 forbids mixing dialects and is a property of a whole text; a single rendering cannot violate it. Both runs that logged an across-text F5 row coded it D, and the survivor count §5 defines is undefined on it. Two of the 39 published D codes are of this kind. Two more are multi-site rows logging decisions at several different source terms at once. Four of 39 published D codes are not decisions between live options at all, and a pooled coverage rate has been treating them as if they were.

Fifth run, and a sixth missing rule — added S077. The rule set above is STILL UNCHANGED and still frozen v1.0

RS-20260801-strata / E-20260801-strata, verifier 130 checks / 0 failures / 6 mutation tests all caught. T-max-havelaar-i-R07-v1, Multatuli, Dutch, 78 sites, both tests at the moment of decision — the longest R07 log the project has, and the fifth consecutive run to declare American English.

9. M6 — nothing in F1–F10 bears on whether the reader knows who a proper name refers to, and the count of missing rules is now six. The Dutch run's second paragraph is an argument with Hieronymus van Alphen, a Dutch children's poet no anglophone reader has heard of. Gloss him, footnote him, or leave the bare name: F4 reaches proper-noun conventions, not proper-noun knowledge, and F10 excludes the footnote as a device the reader would notice. The gloss and the bare name both survive every rule, and the choice that actually decides whether the paragraph is readable is unruled.

10. M3 should be rewritten, and this run says what it is really about. It was recorded as a third-language hole. On this run it fired four times — on two Amsterdam street names, on a diminutive personal name, and on the name of an institution, none of them a third language, and at each of the four F4's heading and F4's clauses give opposite answers. On the evidence now standing, M3 is not about third languages; it is about F4 having a heading ("No foreign matter") that is wider than its own clauses. A v1.1 that fixed the third-language case alone would leave the hole open.

11. The strict test decides where the rule set is about English and stops where it is about the source. Of the run's four strict D codes in 78 sites (5.1%, against an exclusion-test rate of 73.1% — a gap of 67.95 points, replicating S067's 69.1), three turn on F1, F5 or F7 — the three rules that name a formal property of English and say nothing about the source. This is RS-20260730h §2's retrospective finding arriving prospectively.

12. And the second reader §Known limitations calls for does not exist yet. That section says the D/P/S code is assigned by the translator and that "a second reader is a separate, budgeted check". RS-20260730h §3 built one for D; RS-20260801 built one that can express all three, passed all four of its synthetic controls, and then recovered the lead's P and S codes at 0.4118 against a majority-class baseline of 0.4706 — below a constant guess, with a prose-only null scoring 0.5294 on the same items. The mechanism is that three frontier seats differ threefold in how many of the ten rules they judge to bear at all (0.79 to 2.29 per site), and every code statistic in that run is rank-ordered by that number. Any page reporting a cov(R07) figure should now also say that two of its three codes have no second reader.

Binding condition, added 2026-07-31 (S073) — the discharge of a wiki/backlog.md row that reached the review-or-retire rule

The row said R07 needs interpretive rulings its own source does not contain and is silent on five things (opened S063, widened S067, over age at S073). It is discharged here by conversion into a condition at the point of use rather than by another session of the table — the S061 precedent (framework/control-arm-spec.md), taken because a row nobody has to act on is a row that ages.

Any design that translates under R07, or that computes a coverage figure over an R07 log, must settle the following in its own text before it opens the source, and say which reading it took:

• IR1 — F10's scope. Read literally, "nothing that calls attention to the language" forbids carrying over any figure the source contains.
• IR2 — F6's scope. All three published runs read F6 as licensing departure from word order and forbidding paraphrase; F6's text licenses departure and forbids nothing. IR2 turns D codes into P codes wherever it is applied.
• M1–M5 — the five things F1–F10 do not cover, of which M5 (second-person address systems) and M3 (a third language inside the source) are the two known to fire in practice.

And it must not compute the exclusion-test coverage rate at all unless it also reports the mean number of live options per site, because RS-20260730h showed the two are rank-identical across all four runs. A v1.1 of this regime would settle these; nothing currently requires one, and this condition is what stands in until something does.