Repository path: workshop/experiments/E-20260809c-generic-voice/design.md · rendered 2026-09-09

E-20260809c — does genericising a narrator move voice and leave naturalness alone?

ARM-voice-generic step 1. Frozen before dispatch. Nothing below was written after a number existed. The translation and its log were frozen first, at 51d2fcc, before this file was begun.

1. The question

wiki/goodness-senses.md, voice: "Distinguish from naturalness: a natural voice can be a generic voice." Never measured. S135's reachability survey on the style-correspondence entry names the operator: "an operator that genericises a narrator while holding register would move one and not the other."

One direction is not enough. If voice were simply a noisier reading of naturalness, a genericising operator that happened to cost a little fluency would produce a voice drop too, and a single-direction design could not tell that from separation. So this run builds both operators and tests the interaction:

• GENERIC genericises the narrator's manner and should cost voice and not naturalness.
• AWKWARD roughens the English at sites that are not voice carriers and should cost naturalness and not voice.

A double dissociation is also the only form immune to the obvious objection about extent — that an arm which changes more text moves more scales. The primaries are differences between two senses measured on the same arm, so an account in which editing simply moves everything predicts both primaries at zero.

What this unit teaches about translating literature (subject rule, wiki/tracks.md): whether the thing a translator is doing when they try to keep a narrator — as against keeping the propositions and writing well — is visible to a reader as a separate property at all. The alternative outcome, that it is not, is the same finding RS-20260807f reached about perceived-source-carriage, and it would say that this project's controlled typology has two labels for one reader impression.

2. Materials

• Source. Jan Neruda, «O měkkém srdci paní Rusky», Povídky malostranské, 2nd edn (Praha: Valečka, 1885), pp. 107–112; first published 1875. Public domain. 630 Czech words in seven segments (workshop/translations/ruska/source-cs-segments.txt), the story's opening five paragraphs, all of it narration. Single witness, uncollated — declared on the translation artifact. The project's first Czech, its first West Slavic source, its eighteenth source language.
• Why this passage. The narrator is present in the first person twice (Myslím, Měl jsem), disclaims his own gossip twice (prý, prý), announces that a thing is nobody's business and then tells it, keeps two sentences with no verb, and runs sixty words of an imagined scene in the present tense. Those are manner, not content, which is what makes the genericising operator constructible at all.
• CLOSE = T-ruska-R04-v1, frozen at 51d2fcc before this design existed, with its translator's log (D1–D23).
• The three derived arms are built and asserted by materials/arms.py, which runs 6 families of mechanical check and refuses to export an arm that fails one.

The four arms

What holds the two operators apart, mechanically. arms.py asserts that ten named CLOSE strings — the inverted ledger name, the three-verb asyndeton, both verbless sentences, the present-tense soup, the gnomic always say, the fronted object with its two they says, the direct-speech punchline, the asyndetic tricolon — are absent from GENERIC and present verbatim in AWKWARD. GENERIC exists to move exactly those; AWKWARD may not touch one.

What holds content, mechanically. Two checks, not one. (i) Class C: 41 declared strings — every image, every evaluative word, every named entity, every number — must appear byte-identical in CLOSE, GENERIC and AWKWARD. (ii) Content-word conservation: every content word of CLOSE must survive into each derived arm on a 4-character stem match, or be listed in DECLARED_LOSSES with the operator kind that removed it. There are 20 declared losses and every one is a lexical move belonging to a named operator. This is the analogue of RS-20260808d's Class A / Class B split, made stricter: that run held its image-bearing strings identical, this one holds those and audits the whole content vocabulary.

Non-calque discipline. RS-20260807f's pre-run critic caught six of eighteen unlicensed edits that were calques of the source at the site, and RS-20260808d banned fronted inversion outright because Danish is V2. Czech is article-less, case-marked and orders constituents freely, so an English article omission, a bare-noun subject or a fronted object would be a calque here. None of the 21 AWKWARD edits is of those kinds — every one is a wooden habit of English officialese — and each row of AWKWARD_SITES states what the Czech does at that site. Two rows say so explicitly: a leather grocer's apron → an apron of leather for grocers moves away from the Czech, whose shape is adjective + adjective + noun.

3. Panel and stages

Seats J1 = P1 (openai/gpt-5.6-terra), J2 = P2 (google/gemini-3.6-flash), J3 = P5 (deepseek/deepseek-v4-pro) — the S135/S141 seats, unchanged, so the naturalness figures sit on the same panel as the runs they will be compared with. Critic and the parity control are non-panel. The lead judges nothing (charter §5).

28 items (4 arms × 7 segments) per seat at each rating stage; 3 seats; 84 naturalness cells + 84 register cells + 84 voice cells plus 15 competence cells. Item ids are opaque hashes, order is seed-fixed and independently shuffled per seat, and arms are interleaved.

voice is scored alone. Usage rule 6 (S135) requires accuracy to be scored in its own call where a formal comparison is load-bearing, because the same difference measured +0.19 alone and +0.57 alongside three senses in one experiment. The same argument applies with more force here: the whole question is whether voice and naturalness move together, and putting them in one call would manufacture the correlation the run exists to measure. They are in different calls, in different stages, and naturalness is answered before any seat has seen a word of Czech.

The two required quote fields at stage 4 are note (bky)'s remedy, made a first-class measure: each voice rating must name (i) the English phrase most responsible for it and (ii) the Czech it answers to. Gate G6 reads the first and gate G1b reads the second, and both are checked before the primary.

4. Gates — declared here, none movable after a number exists

There is deliberately no gate requiring GENERIC to move voice. That is P1, and a run may not gate its primary on itself.

5. Predictions, registered

Read on seat-pooled segment means, n = 7, with the 21-cell form reported alongside. Exact paired permutation over sign flips, 2⁷ = 128 relabelings, minimum attainable P = 0.0078. Δnat is from stage 1, Δvoice from stage 4; any constant offset between the two stages cancels inside each difference, which is why the primaries are differences of differences.

• P1 — the forward dissociation. D_G = Δvoice(CLOSE−GENERIC) − Δnaturalness(CLOSE−GENERIC) ≥ +1.00, exact P ≤ 0.05. Holds ⇒ genericising the manner costs the person-sense and not the fluency-sense.
• P2 — the reverse dissociation. D_A = Δnaturalness(CLOSE−AWKWARD) − Δvoice(CLOSE−AWKWARD) ≥ +1.00, exact P ≤ 0.05. Holds ⇒ roughening the English costs the fluency-sense and not the person-sense.
• P3 — the parenthesis in its own words. "A natural voice can be a generic voice": Δnaturalness(CLOSE−GENERIC) ≤ +0.25 (amended at §9 A4; as first written this was "≤ 0 or the interval contains 0", which was close to unfalsifiable). Registered as a directional claim and not as an equivalence test, because RS-20260808c and RS-20260809b have now both shown that an equivalence interval on seven segments cannot pass at this n (note (bkm)); an equivalence claim here would be unfalsifiable-by-arithmetic and is not made.
• P4 — the crossing. GENERIC scores higher than AWKWARD on naturalness in 7 of 7 segments and lower on voice in 7 of 7 (amended at §9 A5: 6 of 7 gives an exact sign-test P of 0.0625, which does not reach 0.05). P4 requires both halves; the counts are reported whatever they are.
• P5 — reported, not predicted. r(blind naturalness, voice) over the 28 items. If voice were naturalness under another name, this is where it shows: RS-20260807f measured −0.975 for the last sense that turned out to be one.

6. Failure criteria — declared, and not weakened after firing

• F1. G1a fails for ≥ 2 seats, or G1b < 0.80 ⇒ stage 4 is uninterpretable; P1, P2, P4 WITHHELD.
• F2. G2 fails ⇒ the genericising operator did not hold register; P1 and P3 WITHHELD, P2 unaffected (it does not use GENERIC).
• F3. G3a < 6 of 7, G3b < 6 of 7, or G3c < 4 of 5 ⇒ the arms are not content-matched; everything in §5 is descriptive and no sentence enters wiki/goodness-senses.md.
• F4. G4 fails ⇒ P2 and P4 WITHHELD.
• F5. G6 fails ⇒ P1 is reported as descriptive only and may not be written into the typology as a measured separation.
• F6. Any seat returning < 80% of its cells is dropped and the loss is declared; cell completeness is reported whatever it is.
• F7. If P1 holds and P2 fails, the result is reported as a single dissociation, which is compatible with voice being a noisier naturalness, in those words, and the typology sentence says so.

7. Known limits, written before the run

1. The lead wrote all four arms, chose the passage, and wrote the operator tables. G3a/G3b/G3c put content parity outside the lead and G2 puts the register-hold clause outside it; nothing puts the two operator tables themselves outside it. A reader who thinks GENERIC also removed something the lead did not classify as manner has materials/arms.py and the site tables to check, and no third party checked them before dispatch except the stage-0 critic.
2. One passage, one language pair, one narrator. Neruda's narrator is unusually overt. A result here does not establish that voice separates from naturalness on a covert narrator, and the result page will not say it does.
3. AWKWARD is a made object. Real unnatural translations are unnatural for reasons — calque, over-literalism, register error — and this arm is unnatural for no reason, exactly as RS-20260807f's and RS-20260808d's STRANGE arms were. Its nulls are about manufactured woodenness.
4. Note (bkz) does not bind and the reason is stated rather than assumed. That note requires a non-lead arm whenever a finding is about what a regime does. This finding is about what two operators applied to one text do to two scales; there is no regime in it, and all four arms are derived from one rendering, so a lead-specific translating habit sits identically in every arm and cannot produce a difference between them. What the note would bite on, and does: any sentence claiming that translators who genericise lose voice — no such sentence is registered here.
5. contamination is suspected and unmeasured, with no reachable comparator: both English Povídky malostranské collections are in copyright. Declared on the translation artifact. The design does not turn on it (limit 4's reasoning).
6. GENERIC edits 18% of CLOSE's tokens and AWKWARD 12%. G0 bounds the ratio and the interaction form makes extent non-explanatory for the primaries, but it remains true that the two operators are not the same size, and any figure compared across arms rather than across senses carries that.
7. Tier D is NOT PASSED. No jury verdict here carries evidential weight; every evaluative sentence in the result is provisional and internal-judgment-only.

8. Pre-flight cost estimate

Built from max_tokens, not from expected length (note (abc)).

Declared worst case: $1.20. UTC day 2026-08-09 stands at $1.460842 of $5.00 before this session, so the headroom is $3.539158 and this run fits it three times over.

9. Amendments after the pre-run critic, before any rating call

Critic: nvidia/nemotron-3-ultra-550b-a55b, reasoning off per note (bkw), $0.0185328, one call, runs/critic.txt. Verdict NEEDS-REDESIGN, on three findings marked BLOCKING. All three are factually false, and they are overruled with the demonstration rather than with an argument.

Overruled — BLOCKING 1 and 2, "GENERIC drops the Class C string the shavings, because it appears there as of the shavings." The check is a substring test and "the shavings" is a substring of "of the shavings". python3 arms.py exits 0 on that string and did so before the critic ran. The critic simulated the code instead of the semantics of in.

Overruled — BLOCKING 3, "AWKWARD edits the declared voice carrier a leather grocer's apron (line 312)." That string is not in the carriers list, at line 312 or anywhere. It is in AWKWARD_SITES and nowhere else, which is exactly where a non-carrier edit belongs.

Overruled — BLOCKING 4/verdict item 3, "carriers names a string GENERIC does not edit (her face, her lips began to tremble, copious tears sprang out), and GENERIC has no S7 edit." GENERIC_SITES has three S7 rows, one of them at precisely that string, and the assertion c not in joined_gen passes for it.

Six findings are accepted, and one of the false ones pointed at a real weakness.

• A1 (from the false BLOCKING 1). The Class C check ran on the seven segments joined, so a string could in principle have been satisfied by a coincidental match in a different segment. It is now checked per segment, against the segment of CLOSE the string actually comes from. All 41 strings still pass.
• A2. A kind-disjointness assertion is added: the GENERIC kind set {RH, DT, ID, DI} and the AWKWARD kind set {NOM, PAS, PAD, PER} must not intersect, and no kind outside those may appear. overlaps() additionally reports three spans both operators touch — S2 every comer, S4 appointed for four, S6 and once they beat him, each a DT edit in one arm and a NOM edit in the other. Site overlap is declared, not forbidden: the design's claim is disjointness in kind, the arms are independent derivations from CLOSE, and a span edited by two different kinds in two different arms is not a confound between them.
• A3 (critic SERIOUS b, accepted). G6's bar of ≥ 0.33 was near chance. The chance baseline is now computed and committed before dispatch: 0.2887 — the share of GENERIC's tokens lying inside a declared edit span, arms.py: edit_span_share(). G6's bar is raised to ≥ 0.50 and must additionally exceed 0.2887.
• A4 (critic SERIOUS c, accepted). P3 as registered ("≤ 0, or its 90% interval contains 0") was close to unfalsifiable, which is the defect it was written to avoid. P3 becomes a plain directional bar: Δnaturalness(CLOSE−GENERIC) ≤ +0.25. The interval is reported as description and decides nothing.
• A5 (critic SERIOUS d, accepted — the arithmetic is right). At 6 of 7 an exact sign test gives P = 0.0625, which does not reach 0.05, so P4 could have been called at a significance it does not have. P4 now requires 7 of 7 on both halves. The observed counts are reported whatever they are.
• A6 (critic SERIOUS a, overruled on substance, wording fixed). The critic read G2 as having 21 observations against the primary's 7. Both are 7 segment means, each pooled over 3 seats, 21 ratings per arm behind them — the precisions are equal, which is what note (bkv) asks. The sentence in §4 is reworded to say both numbers so the ambiguity cannot recur.
• A7 (critic MINOR, accepted). Two no-op rows removed from DECLARED_LOSSES and eight dead single-character entries removed from STOP (the checker already skips tokens shorter than three characters).

The critic's power paragraph is accepted as a limit and not as an amendment. It is right that P1's and P2's conjunction of ≥ 1.00 with exact P ≤ 0.05 is demanding at n = 7; it is wrong that power is "near zero", because the sign-flip test's P depends on the consistency of the seven differences and not on their size — seven same-signed differences give P = 0.0156 whatever their magnitude. What is true, and now travels as limit 8: the ≥ 1.00 bar is the binding half of each primary, and a real separation smaller than a scale point fails these predictions as registered.