# Inherited, Self-Built, or None: Does a Translator-Persona's Provenance Shape a Literary Translation?

### A four-way case study of one chapter of Gabriele Tergit's *Effingers* — three AI renderings that differ only in how their translating persona was formed, read alongside the established human translation.

> **Stance and provenance.** This paper is written by the *informed* (non-blind) analyst arm of the
> study. It is built from a prior close reading (`analysis.md`) and the working notes in `notes/`, and
> every claim is grounded in quoted evidence from the source chapter, the four translations, and — for
> the three AI arms — their personas and pass logs. Throughout, the four renderings are treated as
> **co-equal columns**. The work is **description, clustering, and evidence-grounded analysis, not a
> ranking.** I characterize closeness to the German and patterns of divergence, but *which rendering is
> the better or more aesthetically successful literary translation is a judgment reserved for human
> reviewers and is deliberately not made here.* A separate **blind** comparison of the same four texts
> was run independently and withheld from me, for later triangulation.

---

## Abstract

We ask whether the **source** of an AI translator-persona's "soul" shapes the resulting literary
translation, and how AI renderings stand in relation to an established human translation. Four English
versions of one German chapter — Chapter 25 ("Frühling") of Gabriele Tergit's novel *Effingers* — are
compared as co-equal columns. Three are AI renderings that differ *only* in how the translating persona
was formed: **A** built its persona from an existing body of interpretive and translation-craft writing
*about* the author, by the established human translator of her work (an *inherited* frame); **B**
self-authored its persona from a 46-document corpus it assembled (the author's works, scholarship,
contemporaries), excluding the human translator's writing (a *self-made* frame), and revised that
persona after reading the full novel; **C** is the control (no persona, no research). The fourth, **H**,
is the human translator's own published version. Reading all four against the German and coding 27
discrete divergence loci, we find that the three AI arms **cluster tightly** (mean pairwise agreement
≈ 73% of loci) while **H stands apart** (mean agreement with the AI arms ≈ 21%). Within the AI cluster,
**B and C are the closest pair** (81%) and **A is the most distinct**. The expectation that A — whose
persona was built from H's own writing — would resemble H more than B or C do is, on the whole, *not*
borne out: A is marginally the closest AI arm to H (8 of 27 loci vs 4–5), but the margin rests on only
two loci where A sides with H against both other AI arms, and both are discrete lexical rules A imported
verbatim from the human translator's essays; on the dominant structural and stylistic features A clusters
with B and C. Because the persona-free control reaches the same cluster, we attribute the convergence
chiefly to the **shared base model and the shared full-read protocol**, not to the personas, which differ
greatly in *conception* yet leave a *modest footprint* on the output. We report process/effort,
documented reasoning, and the persona contrasts. The study is an **n = 1** qualitative case study with an
informed analyst and a base model shared across the AI arms; findings are descriptive, not inferential.

---

## 1. Background and research question

Literary translation is often described as the production of a *voice* — not chiefly what is said but
how it is said. If a translating agent's voice is shaped by a constructed authorial "persona," a natural
question is whether the **provenance** of that persona — where the agent's understanding of the author
comes from — leaves a detectable mark on the finished translation, and how any such mark compares to the
choices a human translator actually made.

The present case study isolates persona provenance while holding the source text, the procedure, and the
base model fixed across three AI arms, and adds the established human translation as a fourth, co-equal
column. The research question, as posed by the study's coordinator, is twofold:

1. Does the *source* of a translator-persona's soul shape the resulting literary translation?
2. How do the AI renderings stand in relation to the established human translation — and in particular,
   does the arm whose persona was built from the human translator's *own writing about the author*
   resemble that translator's published version more than the other arms do?

The text under translation is Chapter 25 of *Effingers*, titled "Frühling" ("Spring"). It is a montage
chapter: one Saturday — 16 March 1887 — narrated hour by hour, cross-cutting the whole society (the
wealthy Eugenie and her seamstress; the infatuated girl Sofie; the Jewish jurist Waldemar and the singer
Susanna; the working-class Chausseestraße; the jilted Theodor and the child-prostitute Wanda), stitched
together by a recurring time-stamped **spring-refrain**. Its translation-bearing features — the refrain
and its deliberate variants; the montage paragraphing; Berlin working-class dialect; quoted song and
opera titles; forms of address and proper nouns; a high legal-political register in the Waldemar scene;
and Tergit's terse, verbless, paratactic sentence-rhythm — give many discrete sites at which translators
must choose, and so make the chapter a useful probe.

---

## 2. Methods

### 2.1 Design and the four renderings

A single German chapter was rendered into English four times. Three renderings (A, B, C) were produced
by AI translators under controlled conditions differing *only* in persona provenance; the fourth (H) is
the human anchor.

- **A — inherited frame.** A constructed a first-person authorial persona ("the self I translate from")
  from an existing body of critical, biographical, and translation-craft writing *about* the author, by
  the established human translator of her work. A's own step-2 notes name the source: the translator is
  **Sophie Duvernoy** (*Käsebier*, NYRB 2019; *Effingers*, NYRB), and A built the persona from four
  Duvernoy documents, including Duvernoy's introduction to *Effingers* and a blog essay that works
  through a sample translation of the same novel ([A_step2 L12–17](inputs/agent_a/A_step2_persona_notes.md)).
  Thus H is not merely "a human anchor": **H is the very translator whose writing became A's persona** —
  which makes the A-vs-H comparison (§3.5) the study's sharpest test.
- **B — self-made frame.** B assembled and read a broad corpus itself — the author's primary works,
  biographies and scholarship, and her contemporaries and literary tradition — and *self-authored* its
  persona from that corpus. By design the corpus **excluded any writing by the human translator** whose
  frame defines A (control verified: zero mentions of Duvernoy across B's files). B additionally
  **revised its persona after reading the full novel**; both versions (before/after) are available.
- **C — control.** No persona, no research beyond reading the novel: C read and translated.
- **H — human anchor.** The established human translator's own published English version of the chapter.
  Only the finished text exists — no persona, construction notes, or pass log — so H's approach is
  inferred from the text alone, and labelled as inference wherever it is drawn.

### 2.2 Procedure, controls, and constraints

The three AI arms followed a four-step procedure (A and B did steps 1–4; C joined at step 3): (1) read
the author's primary corpus (5 works), notes only; (2) constitute the persona, notes only; (3) read the
full novel (*Effingers*, 151 chapters plus an epilogue), notes only; (4) translate Chapter 25 over up to
five logged revision passes, each agent deciding how many to take. Controls: all three AI arms had **no
internet access** and **minimal system prompts** (no style, epoch, or usage guidance — the interpretation
was the agent's own); a **full-read protocol** was enforced at steps 1 and 3 (read in full, no sampling),
using fresh instances after a pilot in which an instance economized by sampling; a uniform **five-pass
cap** at step 4; and verification that arm B never saw the human translator's writing. H is a published
human translation and did not undergo this procedure. **The three AI arms share a common base model** — a
point that bears directly on the interpretation of their convergence (§4).

### 2.3 Quantitative measurements

Supplied by the coordinator (facts and measurements only):

| | A | B | C | H |
|---|---|---|---|---|
| Persona source | inherited frame | self-built (46-doc corpus) | none | — (human) |
| Step-1 corpus notes | ~58,000 w | ~48,000 w | — | — |
| Step-3 novel notes | ~21,400 w | ~48,800 w | ~13,600 w | — |
| Persona length (after step 2) | 2,257 w | 2,425 w | — | — |
| Persona revised after the novel? | no | **yes (+~980 chars)** | — | — |
| Translation revision passes | 2 | 4 (reached the 5-cap) | 3 | — |
| Final translation length | 4,294 w | 4,454 w | 4,418 w | ≈4,285 w |

### 2.4 Analytic stance and coding

I read the German source in full (extracted from the supplied `.docx`), then all four translations in
full, then — for the AI arms — every persona, construction note, reading note, and pass log. I assembled
the chapter's salient features and, for each, placed the German beside all four English renderings,
recording *that*, *where*, and the *nature* of each difference. To convert this qualitative reading into
a transparent, reproducible clustering measure, I coded **27 discrete divergence loci** (Appendix A),
each by the *salient strategic choice* it presents, and partitioned the four renderings into agreement
groups at each locus. A pair of renderings "agrees" at a locus if both fall in the same group; pairwise
agreement is the count of loci (out of 27) at which a pair agrees. This is an explicitly **descriptive**
instrument — one defensible coding among several, presented for transparency — and **not** an inferential
test; with n = 1 chapter no statistical claim is made. For H, which has no process record, every
statement about *reasoning* is an inference from the finished text and is marked as such.

A note on the word *fidelity*. Where I describe a rendering as tracking the German construction or
lexis closely (or, conversely, recasting freely), this is a **descriptive** observation about
source-closeness, not a quality verdict. Closeness to the source is not equivalent to literary merit; a
freer rendering may be the superior literary translation. That judgment is reserved for human reviewers.

---

## 3. Results

### 3.1 Process and effort

All three AI arms logged their passes; H has none. The logged pass counts reconcile with the
measurements as *one draft plus revisions*: A logged three entries (a draft plus two revision passes,
the second making no changes → "2 revisions"); B logged five (a draft plus four revisions, reaching the
five-pass cap); C logged four (a draft plus three revisions, holding the fifth pass in reserve).

The procedure functioned **differently** across arms. For **B and C the multi-pass process caught real
mistranslations**: B's pass 2 corrected two outright errors in its own draft — *vergessen* rendered
"forgive" → "**forget**," and the Gotthard-Tunnel idiom whose agency it had reversed (*"hat mich auf dem
Gewissen"*, "has me on its conscience") — and C's pass 4 caught a directional idiom, *entgegenlachen*
"laughing back at us" → "**laughs to greet us**," and earlier self-corrected a Wagner calque ("Fire
Magic" → "Magic Fire Music"). **A's revisions found no outright meaning-errors**; its pass 2 made five
small de-*translationese* fixes and its pass 3 changed nothing. Stopping behaviour also differed: **A
stopped early "by choice,"** invoking "the lightest possible hand"; **B went to the cap**, audibly
weighing whether to stop ("to refuse even to look would have been pride dressed up as principle"); **C
held one pass in reserve.** Effort upstream varied widely too (Table, §2.3): B's step-3 novel notes
(~48,800 w) were more than double A's (~21,400 w) and over triple C's (~13,600 w), yet — as §3.2 shows —
upstream effort did not translate into proportionate divergence of output.

### 3.2 Where the four diverge and converge — the clustering

**Headline.** On nearly every pervasive, load-bearing feature, **A, B, and C cluster together and H
stands apart.** The coded tally (Appendix A) makes the shape explicit: pairwise agreement among the AI
arms is **AB = 18, AC = 19, BC = 22** (of 27 loci; mean ≈ 73%), while agreement between each AI arm and
H is **AH = 8, BH = 4, CH = 5** (mean ≈ 21%). Within the AI cluster, **B and C are the tightest pair**
(22/27, frequently word-for-word identical), and **A is the most distinct of the three**, chiefly because
it retains more German and makes a handful of its own lexical calls. A small number of *cross-cutting*
splits divide the field otherwise (§3.2.8). The feature-by-feature evidence follows.

#### 3.2.1 The spring-refrain
German (10 a.m.): *"Was für ein Frühlingstag, dieser Sonnabend im März des Jahres 1887! Was für eine Süße, morgens um zehn Uhr!"*
- **A:** "What a spring day, this Saturday in March of the year 1887! What sweetness, at ten in the morning!"
- **B:** "What a spring day, this Saturday in March of the year 1887! What sweetness, at ten in the morning!" *(verbatim A)*
- **C:** "What a spring day, this Saturday in March of the year 1887! What sweetness, at ten o'clock in the morning!"
- **H:** "What a beautiful spring day, that Saturday in March 1887! How sweet the air was at ten o'clock in the morning!"

The AI arms converge almost exactly — literal *Süße* → "**What sweetness**," the demonstrative, and "**of
the year** 1887" (A and B verbatim; C adds only "o'clock"). H diverges on three counts at once: it adds
"**beautiful**," renders the noun-exclamation as the clause "**How sweet the air was**" (supplying "the
air" and a finite past-tense verb), and drops "of the year." All four nonetheless preserve Tergit's
*variants* of the refrain (the 6 p.m. instance dropping "of the year" and swapping in a second
exclamation; the 8 p.m. single clause; the 3 a.m. return to the full form), so the *tracking* of variants
is four-way convergent even as the *wording* splits A/B/C vs H.

#### 3.2.2 The montage — paragraphing of the refrain
This is H's single largest structural departure, and it is verifiable in the raw text. In the source and
in all three AI versions each refrain stands **alone**, set off by blank lines — the white-space beat that
makes the chapter read as a clock-montage. H instead **runs every refrain into the adjacent paragraph**,
joining the end of one scene, the refrain, and the start of the next into one block:

> **H ([L73](inputs/human/H_translation.md)):** "…will you take the girl out for a walk through the
> Tiergarten later?" What a beautiful spring day… How sweet the air was at one o'clock in the afternoon!
> It was time for the changing of the guard. Standing at the window of his university office, Waldemar
> Goldschmidt, lecturer, gazed out…"

> **A ([L83–85](inputs/agent_a/A_translation.md)):** … ⏎⏎ "What a spring day, this Saturday in March of
> the year 1887! What sweetness, at one in the afternoon!" ⏎⏎ "The guard was changing! At a window of the
> university stood the Privatdozent Waldemar Goldschmidt…"

The fusion is consistent across all the chapter's refrains, so it reads as an intentional feature of H's
continuous-prose style rather than a stray artifact — and it dissolves precisely the montage spacing the
AI arms keep.

#### 3.2.3 Berlin dialect
German (Chausseestraße): *"»Paule, muß denn det sin?« … »Wird er doch wieder'n Wochenlohn versaufen!«"*
- **A:** "Paule, must that be?" … "He'll go and drink up a week's wages again!"
- **B:** "Paule, does it have to be, then?" … "He'll go and drink up a week's wages again!"
- **C:** "Paule, does it have to be, then?" … "He'll go and drink away a week's wages again!"
- **H:** "Oh Paulie, d'you haff to?" … "You'll piss away your week's pay again!"

A clear spectrum. **H carries class through heavy eye-dialect** (phonetic respelling — "Paulie," "d'you
haff to," elsewhere "leggo," "yer," "Can'tcha") and a blunt colloquial-American lexis ("piss away,"
"Stupid cow," "What a sop!"). **A and B carry class through diction and idiom**, with very light
eye-dialect and a British flavour ("leave off," "daft goose/goat," "carry on so"); **C sits a touch
rougher** (double negatives, "Aw"/"Naw," "'cause") but still avoids phonetic respelling. Notably, all
three AI arms articulated the *same* governing principle independently — render Berlinisch as generic
working-class English, **not** Cockney (A: "NOT Cockney/Brooklyn"; B: "I will not turn Berlinisch into
Cockney … that is the translator's vanity"; C: "did NOT map onto a specific English dialect … to avoid
importing a foreign locale"). The closing *"Der Dämel!"* is a four-way divergence with no two alike: A
"The chump!", B "The great fool!", C "The ninny!", H "What a sop!"

#### 3.2.4 Quoted song and opera titles

| | Schumann cycle | Wagner piece | Heine/Schumann incipit |
|---|---|---|---|
| A | *A Woman's Love and Life* (translated) | the "Magic Fire" | "The lotus flower is afraid…" |
| B | 'A Woman's Love and Life' (translated) | the "Magic Fire Music" | "The lotus flower fears…" |
| C | *A Woman's Love and Life* (translated) | the "Magic Fire Music" | "The lotus flower fears…" |
| H | "**Frauenliebe und Leben**" (German) | Wagner's "**Feuerzauber**" (German) | "The lotus flower **trembles**…" |

Opposite surface strategies: the AI arms **translate** the German titles into English; **H keeps them in
German** and, where the AI arms strip apparatus, **adds the attribution "Wagner's"** not present in the
source. The Schumann lyric is rendered plainly by all three AI arms (A: "Since I beheld you, I think
myself blind; ever, as in a dream, I see you alone"); **H alone makes a rhyme** ("…I believe I've gone
blind, as if in a dream, only you in my mind"). And **B alone** capitalizes "You/Your" through Sofie's
letter to mirror the German reverent *Du*, a choice B documents ("to carry the intimacy of the German
*Du*-letter").

#### 3.2.5 Forms of address and proper nouns (the foreignization gradient)
- **die Widerklee:** A/B/C keep the German calque "**The** Widerklee"; **H drops it.**
- **Privatdozent:** A/B/C keep the German title; **H translates** → "lecturer."
- **Street names:** A/B/C keep German + ß (Chausseestraße…); **H anglicizes** to "-strasse." *Weidendammer
  Brücke:* **A keeps the German**; B/C/H → "Weidendamm Bridge."
- **Graues Kloster:** **A keeps the German**; B/C "Grey Cloister"; H "Gray Cloister."
- **Einverstanden?:** A/B/C "Agreed?"; **H "D'accord?"** (a French substitution).

The net foreignization gradient on proper nouns runs **A > C ≈ B ≫ H** — A retains the most German, H the
least. On these features A is the *furthest* of the four from H, complicating any simple A≈H expectation.

#### 3.2.6 The high register (Waldemar / the historian)
German: *"Prämie auf Charakterlosigkeit."* / *"Ochsen, im wahren Sinne…"* / *"Schäm dich, daß du dich schämst!"*
- **A/B/C** (near-identical): "A premium on want of character." / "Oxen, in the true sense[ of the word]…" / "Be ashamed that you are ashamed!"
- **H:** "They've put a premium on men who lack principles." / "**What numbskulls.** That's why these fools worship young girls." / "Shame on you for being ashamed!"

The AI arms render the aphorisms compactly and close to the German; H expands and idiomatizes (abstract
*Charakterlosigkeit* → concrete "men who lack principles"; "Oxen" → "What numbskulls"). The same shows in
syntax: where the German inverts (*"…lagert die Wolle"*), A/B/C keep the inversion ("the wool lies
stored"); **H regularizes to plain SVO** ("The wool was stored in the old palace built by the Margrave of
Brandenburg").

#### 3.2.7 The cluster map
- **B ≈ C** are the tightest pair (22/27) — frequently verbatim ("casino gown," "a martial sound," "Magic
  Fire Music," "The lotus flower fears," "mutton-legs," "the Fresh Mutton," "coupé," "daft goat," "Grey
  Cloister"). With a shared base model and the persona-free control in this pair, this is the base-model
  default in plainest form.
- **A** joins the *strategy* (foreignizing, literal, montage-holding, "What sweetness," "Agreed," "The
  Widerklee," English song-titles, the compact aphorisms) but stands apart in *execution*: it keeps the
  most German proper nouns and makes its own calls — "dinner gown" (vs B/C "casino gown"), "Magic Fire"
  (vs "…Music"), "warlike sound" (vs "martial"), "carriage"/"stool" (domesticating where B/C keep
  "coupé"/"tabouret").
- **H** is the outlier on the dominant axis: domesticating, montage-fusing, eye-dialect, "How sweet the
  air was," German song-titles retained but the lyric rhymed, "lecturer," anglicized streets, French
  "D'accord," American idiom.

#### 3.2.8 Cross-cutting splits
A few loci divide the field otherwise, and they matter for §3.5:
- **A + H vs B + C** on *"Das Recht?"* → A and H "**Justice?**"; B and C "**The law?**" (the one prominent
  place A sides with H against *both* other AI arms; documented as a deliberate, opposite call — §3.3).
  (On the next phrase the field regroups: A, C, and H all write "Roman law," B alone "the law of the
  Romans.")
- **A + H vs B + C** on *coupé* → A and H "carriage"; B and C "coupé."
- **A + C + H** keep "Annettchen"; **B alone** translates → "dear little Annette."
- **A + C + H** use American spellings ("honor," "theater"); **B** is consistently British ("honour,"
  "theatre"). H is uniformly American throughout (also "gray," "kerosene," "pants," "percent"); A and C
  are *mixed* (American spelling, British idiom) — so orthography does not cleanly separate the AI arms
  from H.
- **Singletons:** C alone keeps "*Molle*" (the Berlin word for a beer) and alone renders "Schloss" →
  "Palace"; A alone keeps "Graues Kloster"/"Frischer Hammel." Two four-way splits where everyone differs:
  *Petroleum* → "lamp oil"/"paraffin"/"petroleum"/"kerosene" (A/B/C/H); *Der Dämel* → "chump"/"great
  fool"/"ninny"/"sop."

### 3.3 The AI arms' documented choices and reasoning

A shared, whole-novel-grounded decision underlies the refrain. **All three AI arms independently
discovered, through the full read, that the refrain is not local to Chapter 25** but recurs across the
novel — Ch 68 ("Frühling," 1913), Ch 131 ("Frühling 1930"), and the close in 1948 — and all three
therefore fixed it as a recitable formula built to recur verbatim: A's pass log ("flexes … for the
novel-wide frame (Ch 68 1913, Ch 131 1930, Epilog 1948)"); B's reading notes ("Ch 68 … **THE STRUCTURAL
TWIN OF MY TRANSLATION TARGET … STUDY THIS HARDEST**"); C's pass log ("this exact formula recurs in Ch68
(1913), Ch131 (1930), and **returns as the novel's final words over ruined 1948 Berlin**"). That the
*control* reached this purely from the full read is the strongest single sign that the convergence is a
product of the shared protocol, not the personas.

The richest material for the cross-cutting splits lies in the logs, because two arms identified the *same*
crux and resolved it *oppositely*, each explaining why:

- ***"Das Recht?"*** A kept "**Justice?**," reasoning that the *der Rechte / das Recht* pun is "essentially
  unrecoverable" and "'Justice?' keeps the philosophical lift." B changed its draft "Right?" to "**The
  law?**": "I accept that the pun … goes soft in English; the conceptual spine — a jurist who … thinks at
  once of *law itself* — matters more, and 'law' carries it." Same diagnosis, opposite resolution; the
  field splits A+H vs B+C. (Inference for H, which has no log: from "Justice? … Roman law … can the new
  civil code forego it?" one reads the same "philosophical lift" instinct A articulates.)
- **"Annettchen."** A *kept* the German diminutive as a buried cross-novel clue (its notes: the
  "Annettchen Easter-egg … left buried for the alert reader," linking the seamstress's lost lover to a
  character elsewhere in the novel). C *kept* it for a within-chapter reason ("so the dramatic-irony clue
  survives … Eugenie cannot [connect it]" plus "Eugenie's clean echo"). **B translated it** to "dear
  little Annette," prioritizing the *echo* (making both occurrences identical "so the English reader hears
  the same words repeated back, as the German does"). Three documented rationales; B's lands on a
  different surface.
- **"Kasinotoilette."** B and C both wrote "casino gown" — for *different* reasons (B: "Kasino" =
  officers'-mess; C: the "resort casino … at Nice"). A wrote "dinner gown," its log flagging this as a
  known flattening it considered reversing and did not.
- **coupé → "carriage."** A's log lists this under "domestication of society-words (persona defaults)"; it
  is lifted almost verbatim from Duvernoy's essay ("no reason to retain 'coupé' when 'carriage' does"). B
  and C kept "coupé." A second documented place A's *inherited* frame steers it toward H.

The arms' *deliberative temperament* also differs, as recorded in §3.1: A polished a draft it judged
sound and stopped early; B and C used the multi-pass process partly to catch genuine mistranslations; C
additionally reasoned on the page about parataxis (restoring Käte's coordinate clauses), period-neutral
diction (Mayer's "down to the fact" → "chiefly because"), and anachronism ("clueless" → "unsuspecting"),
all without a persona.

### 3.4 The personas: A vs B, and B before vs after the novel

**Two conceptions of one author.** Both A and B wrote their personas in the first person, *as* Tergit,
and — because both read widely about her — they converge on the biography (Elise Hirschmann, born 1894 in
the Raupachstraße; the two Berlins, poor-east and wealthy-Tiergarten, that formed "the whole of my eye";
the shy court-reporter's "recording-tape" method; the name *Gitter* reversed to Tergit; anti-Zionism held
in tension; the "driven-out"; the deaths of son Peter and husband Heinz; and the governing discipline
that the novel is **"not written from the end"** — no doom, no hindsight). On *who Tergit is*, the
inherited and self-built souls largely agree.

They differ in **kind**. A's persona, drawn from a *translator's* essays, is dense with explicit
**English-craft method**: the Wharton/Powell/Mitford "ear," a "blended, mid-Atlantic, midcentury"
register, "foreign society-words naturalized — a *coupé* becomes a carriage," "honorifics simplified … the
Prussian *Rat* titles all to 'Councilor,'" "Berlinisch with a light hand … not Cockney," bare "said"
tags, "the lightest possible hand." It conceives Tergit *through the eye of someone who has already
carried her into English.* B's persona, drawn from the author's own works, biographies, scholarship, and
contemporaries — and, by design, **no translator's writing** — is dense instead with **moral and
historical sensibility**: "My greatest hatred is the **generalization**"; "I am a historian by training —
Meinecke was my teacher … **Warum, wieso, weshalb Hitler?**"; "I am one of the **Vertriebenen**, the
driven-out"; "**Diagnosis, not prophecy.**" On English tactics it is almost silent, closing: "Quiet.
Exact. Fast. Warm under the cool … Now let me go and see the chapter." This is the cleanest expression of
the design's contrast: **A inherited a translator's blueprint; B inherited (built) an author's interior**
and had to improvise the surface tactics at translation time. (B *independently* reached for one of the
same anglophone touchstones Duvernoy uses — likening a Tergit narrator to "a Mitford ingénue" — a
convergence of literary instinct, not contamination.)

**B before vs after the novel.** B's two persona versions are **identical but for a single added
paragraph** (the "+~980 chars" of the measurements). The addition is a "**second law**": that the true
hero of the book is not a person but **Time** —

> "*der unbarmherzige Motor, der eigentliche Held, ist die Zeit.* Since 1914 we have not lived our lives;
> we have been *lived* … Paul believes he is doing the shoving, and he is being shoved … the current
> takes them, and I hold the reader's eye on the swimmer, never on the flood … the chronicle must *move* —
> scene cut hard against scene, the dated years tolling like a bell."

Its origin is documented: in the full read B encountered Tergit's own statement (*"Der unbarmherzige
Motor und eigentliche Held des Romans ist die ZEIT … wir alle [sind] … seit 1914 gelebt worden"*) and
integrated it. This is a genuine interpretive deepening, and it sits in **tension** with B's own
before-novel individualism ("the single human soul," "individuals first"), which B reconciles ("I neither
punish my people with their fates nor let them master them"). It also runs *opposite* to A's inherited
frame, in which Duvernoy stresses that Tergit *rejected* the title *Der Ewige Strom* ("The Eternal
Stream") precisely so the people would not seem "ineluctably swept up by … history" — "individual choices
matter." So on the agency↔determinism axis, **A's inherited soul foregrounds agency; B's self-built soul,
after the full read, adds the Time-fatalism as co-equal** — a real frame-level divergence, both readings
textually grounded.

**Does the revision surface in the translation? Largely no — and that is explicable.** B's pass-1
governing idea is the *before*-law ("it is not written from the end … the dread is left entirely to the
reader"). Chapter 25 is set in 1887 — the "founding-era idyll," before the post-1914 machinery of Time
grinds — and B's log treats it as the living present where "nobody … knows what 1887 leads to." The "Time
as hero" theme is therefore deliberately *latent* in this chapter; its only faint trace is the "**tolling
bell**" framing of the refrain (pass 3: "the tolling cycle it should"). The before/after change is thus
significant as *interpretation* but leaves little distinguishable signature in B's Chapter 25 *text*
relative to A's or C's — because this is precisely the chapter where the macro-fatalism is held back.

**The gap between A's persona and A's output.** A's persona is the most concrete blueprint of the four,
and A followed it only **selectively, by its own design**. The persona reserves an override: "**I am the
author, not a faithful servant of the author … I keep the last call.** Where my own ear wants to hold a
German cadence that a translator would smooth … I may." A exercised that license toward foreignization. It
*followed* the discrete Duvernoy rules it had verbatim — coupé→"carriage," *Rat*→"the Councilor" (even
Duvernoy's American one-l spelling), "Justice?", bare tags, light no-Cockney dialect — but on the *broad*
strategy it diverged from the inherited domesticating frame: it kept German proper nouns (justified under
a *separate* persona instinct, "Berlin kept present"), translated the song-titles, held the montage seams,
and rendered "What sweetness." Most tellingly, A enacted a montage-preserving principle its persona
*attributes to H's method* ("Hold the seams; the assemblage is the form") that **H's own published text
does not visibly honour** (§3.2.2). The inheritance transmitted Tergit's *interpretation* (genre, tone,
the no-doom thesis, the whole-novel architecture) far more than it reproduced H's *surface* choices.

### 3.5 Where the human (H) sits — including the A≈H question

Treated as its own result: **H stands apart from all three AI arms** on the dominant axis, and it does so
on the features that most define a literary surface — how the refrain sounds and sits on the page, how
dialect is voiced, whether the foreign is kept or naturalized, how quotations are handled, and how
sentences are shaped (§3.2). The coded agreement of each AI arm with H is low and roughly even: **AH = 8,
BH = 4, CH = 5** of 27 loci, against AI–AI agreements of 18–22. Read off the page (H has no notes), H's
profile is a translator working for **continuous, idiomatic, domesticated English**: fusing the montage
into running prose; anglicizing place-names; supplying an added attribution ("Wagner's"); translating
titles to plain English ("lecturer"); reinterpreting detail ("riding hats" for *Schleier um den
Zylinder*); reaching for the freer idiom ("What numbskulls," "famished," French "D'accord") and the rhymed
lyric; and writing dialect in confident eye-dialect. The three AI arms, by contrast, share a more
**literal, foreignizing, source-tracking** default.

**Does A — whose persona was built from H's own writing — resemble H more than B or C do?** A calibrated
answer: **marginally, and not on the whole.** A is nominally the closest AI arm to H (AH = 8 vs BH = 4,
CH = 5), but the margin is small and **far below any AI–AI pair**, and it rests on a handful of discrete
loci. Of A's eight agreements with H, only **two** are cases where A sides with H against *both* other AI
arms — *"Justice?"* and *coupé*→"carriage" — and **both are rules A imported almost verbatim from
Duvernoy's published essays** (its persona source). The other six (keeping "Schloss," keeping
"Annettchen," lowercase *Du*, translating "Molle," the Hammelbeine-as-equivalent strategy, American
spelling) are each shared with at least one *other* AI arm, so they are not A-specific resemblances to H.
Against this short list stands the overwhelming weight of §3.2: on the refrain's wording and layout,
"What sweetness," the translated song-titles, "The Widerklee," the German street-names, "Privatdozent,"
"Agreed," the compact aphorisms, the British-flavoured idiom, and the held montage, **A clusters with B
and C against H** — and on proper nouns A is the *most foreignizing of all four*, i.e. in that dimension
*further* from domesticating-H than B or C are.

The most economical account of why is supplied by the control. **C — no persona, no research — lands
squarely in the AI cluster** (CH = 5; AC = 19, BC = 22), reaching the cross-novel refrain decision, the
no-Cockney principle, the "Annettchen"-as-clue choice, register-differentiation, and leitword consistency
purely from the full read and base-model competence. If the persona-free arm already produces the cluster,
the cluster is chiefly a product of the **shared base model and the shared full-read protocol**, not of
the personas.

---

## 4. Discussion

Three findings stand out.

**(1) Persona provenance shaped *conception* far more than *output*.** The two constructed personas differ
profoundly in kind — A a translator's craft-blueprint, B an author's moral-historical interior — and B's
post-novel revision adds a genuinely new interpretive thesis (Time as the book's hero). Yet the three AI
translations converge tightly (mean AI–AI agreement ≈ 73% of coded loci), and the persona-free control
sits inside that convergence. The personas' detectable footprint on *this chapter's* output is therefore
**modest**: a handful of specific lexical calls (A's coupé→carriage and "dinner gown"; B's "dear little
Annette," British spelling, and capitalized *Du*; C's lone "*Molle*") and, more richly, a differentiated
**vocabulary of documented reasoning**. The strongest determinant of the shared surface appears to be the
two things held constant across the AI arms — the **base model** and the **full-read protocol** — the
latter visibly responsible for the convergent, whole-novel-grounded refrain decision that even the control
reached.

**(2) The human translation is the genuine outlier.** What separates one of these four columns from the
other three is not which soul stood behind it but that **H was produced by a human translator** whose
instinct, on this chapter, was to **domesticate** — to naturalize the foreign, regularize the syntax,
voice dialect in eye-dialect, gloss and reinterpret, and dissolve the montage into continuous prose —
where the AI arms, severally and together, stayed close to the German. We frame this descriptively: the AI
arms track the source's wording, structure, and foreign texture more closely; H recasts more freely into
fluent target-language prose. We do **not** read this as a quality difference. Source-closeness is not
literary merit, and a freer, more domesticated rendering may well be the more accomplished literary
translation; that judgment is reserved for human reviewers.

**(3) The headline expectation is not borne out — with a calibrated caveat.** The design invites the
hypothesis that A, built from H's own writing, should resemble H. The texts show that the inheritance
transmitted H's *understanding of the author* (which A in fact shares with B, and in thinner form even
with the persona-free C) rather than H's *translating hand*. A's persona explicitly built in a license to
override its inherited frame ("I keep the last call"), and A exercised it toward a literal, foreignizing,
montage-preserving result — at times enacting principles it *attributed* to H's method but that H's own
text does not display. The one honest qualification is quantitative: A *is* the closest AI arm to H, but
by a thin margin (8 vs 4–5 of 27 loci) traceable to two persona-imported lexical rules, not to a pervasive
stylistic kinship.

Taken together, the case suggests that — for this chapter, this model, and this protocol — *what a
persona changes is how the translator talks about the work and a scatter of local choices, more than the
global character of the prose*, and that a deeply read but persona-free control can match persona-bearing
arms on most structural and stylistic decisions.

## 5. Limitations

- **n = 1 chapter, one instance per AI arm.** Every pattern here is a case-study observation, not a
  measured effect; nothing supports generalization to other chapters, authors, models, or runs.
- **Shared base model.** The three AI arms share one base model, which is a live confound for *every*
  similarity among them; the convergence we attribute to "model + protocol" cannot be separated into its
  parts within this design.
- **Informed (non-blind) analyst.** I knew the arm identities, including the A↔H link — a standing source
  of **expectation bias**. I mitigated it by pre-registering the A≈H comparison as a hypothesis, coding
  divergence transparently (Appendix A), and reporting the negative/calibrated result the evidence
  supports; but the mitigation is procedural, not a substitute for blinding. An independent **blind**
  comparison of the same four texts exists and was withheld from me, for later triangulation.
- **The coding is one defensible instrument.** Locus selection, granularity, and the "salient-choice"
  calls affect the exact agreement counts; the *qualitative* pattern (AI–AI ≫ AI–H; B≈C tightest; AH the
  weak maximum among AI–H) is robust to reasonable recodings, but the integers are not to be over-read.
- **H has no process record.** All statements about H's *reasoning* are inferences from the finished text
  and are labelled as such; only A, B, and C left logs. The montage-fusion in H is read from the supplied
  file; its consistency across refrains is why it is read as intentional, but it is a property of the text
  as given.
- **Description, not evaluation.** Observations about source-closeness, caught mistranslations, or
  stopping behaviour are about *process and pattern*, not verdicts of quality. No ranking is offered or
  implied.

## 6. Conclusion

Across one chapter of *Effingers*, three AI translations differing only in persona provenance cluster
closely with one another while the established human translation stands apart — the human domesticating
where the AI arms stay close to the German. Persona provenance strongly shaped each AI arm's *conception*
of the author and its *documented reasoning*, but left only a modest footprint on the finished prose; a
persona-free control reached most of the same structural and stylistic decisions, pointing to the shared
base model and full-read protocol as the principal sources of convergence. The expectation that the arm
built from the human translator's own writing would most resemble that translator's published version was
not borne out except in two discrete, persona-imported lexical choices. Whether any of these four
renderings is the better or more aesthetically successful literary translation is a question this study
does not address and does not attempt to answer; it is left, as it should be, to human reviewers.

---

## Appendix A — Coded divergence (27 loci) and pairwise agreement

Each locus is coded by its *salient strategic choice*; groups list the renderings sharing that choice. A
pair agrees at a locus if both fall in the same group. Descriptive only (n = 1 chapter; not inferential).

| # | Locus | Groups |
|---|---|---|
| 1 | Base refrain wording ("What sweetness / of the year" vs "How sweet the air was / beautiful") | {A,B,C} \| {H} |
| 2 | Refrain layout (own line vs fused into paragraph) | {A,B,C} \| {H} |
| 3 | *Süße* → "sweetness" vs "sweet the air" | {A,B,C} \| {H} |
| 4 | *die Widerklee* → "The Widerklee" vs drop article | {A,B,C} \| {H} |
| 5 | *Privatdozent* kept vs "lecturer" | {A,B,C} \| {H} |
| 6 | Street names German + ß vs "-strasse" | {A,B,C} \| {H} |
| 7 | *Weidendammer Brücke* kept German vs "Bridge" | {A} \| {B,C,H} |
| 8 | *Graues Kloster* kept German vs translated | {A} \| {B,C,H} |
| 9 | (Königstr.) *Schloss* kept vs "Palace" | {A,B,H} \| {C} |
| 10 | *Annettchen* kept vs "dear little Annette" | {A,C,H} \| {B} |
| 11 | *Einverstanden* → "Agreed" vs "D'accord" | {A,B,C} \| {H} |
| 12 | *coupé* → "carriage" (domesticate) vs "coupé" (keep) | {A,H} \| {B,C} |
| 13 | *Frauenliebe* title → English vs German | {A,B,C} \| {H} |
| 14 | *Feuerzauber* title → English vs German (+"Wagner's") | {A,B,C} \| {H} |
| 15 | Schumann lyric rhymed vs not | {H} \| {A,B,C} |
| 16 | *Du*-letter "You" capitalized vs lowercase | {B} \| {A,C,H} |
| 17 | *Das Recht* → "Justice?" vs "The law?" | {A,H} \| {B,C} |
| 18 | *Prämie auf Charakterlosigkeit* compact vs expanded | {A,B,C} \| {H} |
| 19 | *Ochsen* → "Oxen" vs "numbskulls" | {A,B,C} \| {H} |
| 20 | *Bärenhunger* → "hungry as a bear" vs "famished" | {A,B,C} \| {H} |
| 21 | *Schäm dich…* literal vs "Shame on you" | {A,B,C} \| {H} |
| 22 | Dialect: light eye-dialect vs heavy | {A,B,C} \| {H} |
| 23 | Pub *Frischen Hammel*: German / "Fresh Mutton" / "Young Ram" | {A} \| {B,C} \| {H} |
| 24 | *Molle* kept vs translated | {C} \| {A,B,H} |
| 25 | *Hammelbeine* idiom: literal "mutton-legs" vs English equivalent | {B,C} \| {A,H} |
| 26 | Spelling honor/theater: American vs British | {A,C,H} \| {B} |
| 27 | Verbless/inverted syntax kept vs regularized SVO | {A,B,C} \| {H} |

**Pairwise agreement (loci out of 27):**

|   | A | B | C | H |
|---|---|---|---|---|
| **A** | — | 18 | 19 | **8** |
| **B** | 18 | — | **22** | 4 |
| **C** | 19 | **22** | — | 5 |
| **H** | **8** | 4 | 5 | — |

*Read-off.* AI–AI pairs (AB 18, AC 19, BC 22) → mean ≈ 19.7/27 (~73%); AI–H pairs (AH 8, BH 4, CH 5) →
mean ≈ 5.7/27 (~21%). **BC = 22** is the maximum (B and C are the tightest pair). **AH = 8** is the
maximum among AI–H pairs (A is marginally the closest AI arm to H), but of those 8 loci only **2** (#12
*coupé*, #17 *"Justice?"*) are A-with-H against *both* other AI arms, and both are rules A imported from
the human translator's essays; the other six are shared with ≥1 other AI arm. (Checksum: the six pair
counts sum to 76, the total agreeing-pair-instances across all loci.)