Observations

Each of the six AI arms kept notes throughout — reading Tergit's other books, building its persona, reading the novel — and logged every revision. What follows is from that record: how they read Chapter 25, where they split, and — at most length — what each one cited when it justified a choice.

A reading the arms shared

Set the chapter in front of each arm and, separately, all six found the same thing: the chapter's spring refrain is not local to it. The same lines return across the novel — in 1913, in 1930, over the ruins of 1948 — and each decided the refrain had to read as a fixed, recurring formula. D-let, the broader-corpus arm, put the constraint plainly: the refrain "must survive being deployed four times in this chapter and three more times across the novel … so the English needs to be reproducible."

A — reading notes

The refrain "flexes for all seven hours and for the novel-wide frame"; "What sweetness" is "the single most important choice in the chapter."

B — reading notes

Chapter 68 is "the structural twin of my translation target — study this hardest."

C — pass log

"This exact formula recurs in Ch. 68, Ch. 131, and returns as the novel's final words over ruined 1948 Berlin. So it must read as a fixed, recitable refrain."

C reached this from the bare text — no persona, no research.

The cruxes — what each arm noticed

At a handful of places the German sets a problem no English can fully solve, and the logs catch the arms reading it. What is worth watching is not only which way each arm finally went but what it saw in the German first, and which loss it decided it could bear. The rendered choices are set side by side on the Findings page; here the interest is the thinking that arrives at them.

"Das Recht?"

The arms that stop to reason it out all see the same thing: Waldemar's der Rechte / das Recht turns on a pun — the right man, then law itself — that English cannot hold in one word. They part on which half to keep. A reads the line for its register and decides the abstraction matters more than the precision, keeping "Justice?" because the pun is "essentially unrecoverable" and "'Justice?' keeps the philosophical lift" into "Roman law … the Civil Code." B reads the speaker instead — a jurist drafting the code — and on its second pass overturns its own draft, judging that the literal sense had drifted:

B — pass 2

I accept that the pun … goes soft in English; the conceptual spine — a jurist who sees the crowd cheer and thinks at once of law itself — matters more, and "law" carries it.

C, with no persona, reads only the surrounding text and reaches the same word from the anchor beside it — "kept 'the law' consistently (anchored by 'Roman law'); the pun is a minor, unavoidable loss." D-aim arrives at law too, but from the principle rather than the line: the domesticating reading asks what the contemporary reader can parse, and decides the wordplay simply has to go.

D-aim — pass 3

The domesticating choice is to drop the (untranslatable) German wordplay on Recht and use the topic-word "law" explicitly.

Three arms reach the same English word by three different routes — the speaker, the surrounding text, the principle — and A alone keeps the lift over the precision.

One turn of the word, though, sat outside what any arm weighed. Recht and its kin make a single dense field in German — das Recht (law, a right, justice), der Rechte and die Rechten (the right one; the Right) — and the line stages it over a crowd's enthusiasm: »… für den Rechten, morgen für den Unrechten. Das Recht?« Each arm settled the line on one sense and moved on; none, in its notes or passes, weighed whether the English might hold the field's broader political register in play rather than close it. The surface reading is sound — this is no mistranslation — but it is one of the places where the German keeps an ambivalence open that English, asked to choose, tends to set down.

"Annettchen"

Every arm notices that the diminutive is a planted echo: the cast-off girl quotes the pet name, and Eugenie fastens on exactly that word. The question each weighs is where the English reader will hear the click — inside a kept German word, or only through an English diminutive. A reads it as an Easter-egg and keeps the German diminutive "left buried for the alert reader"; C, reasoning the same way from the bare text, keeps it too — "so the dramatic-irony clue survives"; D-let keeps it because the German diminutive "carries the unwitting cruelty of the overheard scene." The arms that translate it are reasoning from the reader's ear rather than the page. B, on re-reading, makes both occurrences match so the repetition lands:

B — pass 3

The pet name is the twist of the knife … Now both are "dear little Annette", so the English reader hears the same words repeated back, as the German does.

C-let renders it "Annette dear" on the same instinct — "Annettchen in English would jar" — and D-aim "little Annette," reasoning that "the English reader hears the intimacy through 'little'." The split is not over what the word does; it is over whether English can be trusted to overhear it.

coupé

The narrower case shows the same shape in miniature: the arms read one period word and weigh whether it still lives in English. A applies an inherited default and naturalizes it to "carriage" — a move it carried in from the human translator's essays on rendering society-words. The others test the word against period usage and judge it already at home: D-let keeps "coupé" because "it's in English period vocabulary," D-aim as "the English carriage-term of the period," and C and C-let leave it standing on the same reading. The reasoning, not the result, is the tell — one arm trusting a settled habit, the rest re-checking the word against the period and finding it needs no translating.

How each one worked

The pass logs are working records, not verdicts: each arm says what it was trying to do, and why, before it does it. The moments below are characteristic — an arm naming the warrant for a single word, or catching itself in an error and reasoning its way back.

The clearest of them is B, the self-built arm, switching to British spelling late in revision. It is not a house-style tic but a reading of the persona — and B reaches for the life to justify it:

B — pass 4

A finished literary translation must pick one convention and hold it, and British is also true to the persona: Tergit lived in London from 1938 and wrote English as the English do.

Where an arm catches its own mistake, the log shows the reasoning as much as the fix. C, the control working from the bare text with no persona at all, used a deliberately skeptical last pass to reverse a directional error it had carried since the first draft:

C — pass 4

"All of nature laughs to greet us." Entgegenlachen = to laugh/beam toward someone in welcome … English "laughing back at us" carries the wrong vector — it reads as laughing in return or even mockingly, the opposite of the radiant, benign welcome intended.

Not every arm had an error to catch. A, the inherited arm, revised only twice and stopped — "by choice," invoking "the lightest possible hand" — judging on its last read that what remained would be "fussing, not improving." It was polishing a draft it found already sound.

D-aim, the school arm, used the same kind of pass to undo a misreading that had survived three rounds of polishing — and read the fix back against the whole speech it sat in:

D-aim — pass 4

The English "loves oneself" reads as narcissism, which makes Eugenie sound like she is preaching against self-love — the opposite of what she is saying … "When one does the loving oneself" … the meaning is right, and Eugenie's whole speech coheres around the contrast between being the lover and being the loved one.

And where an arm keeps a German word, it says why it kept it. D-let, the broader-corpus arm, held a cluster of diminutives on the grounds that the German carries a tone English cannot:

D-let — pass 1

Annettchen and Paule are affectionate diminutives that carry their tone better in German than in any English diminutive form.

How the six reasoned, and what each cited

Reading the same passage, six arms reach for different warrants. Coding every justified choice in the six pass logs against the authority it appeals to — the German source, the English reader's ear, the persona, a named anglophone voice, the period, the brief — sorts the arms into distinct profiles. The clearest finding is not which authority an arm prefers, but that the same instruction, worded two ways, is cited utterly differently.

A permission is quiet; a school is loud

D-let and D-aim, the two broader-corpus forks, were run twice from the same point. Both share one persona, one reading of the novel, one instance — they part only at the moment of translation, where D-let carried a quiet permission to domesticate in its brief and D-aim a loud school instruction to do so, the same licence in content. The logs of the two read nothing alike.

The word domesticate, in any form, appears not once in D-let's log and fifteen times in D-aim's. D-let acts on its licence in silence — turning gnädige Frau into "ma'am," Frühstück into "lunch" — and where it pauses to give a reason, it cites the German or the persona, not the brief; elsewhere it keeps the German on source-text grounds. The single term Privatdozent shows the gap whole: the two forks reason from the same shared persona to opposite ends, and each says why. D-let holds the word:

D-let — pass 1

Kept Privatdozent (no English equivalent), Herr Kollege, Kaiserreich, Monumenta, Junker … German social-historical terms with no clean English equivalents.

D-aim, given the school instruction at the translation step, anglicizes the same word — and overrides the very fact D-let used to keep it:

D-aim — pass 1

Privatdozent → the young lecturer … anglicising fully — the contemporary anglophone reader does not parse "Privatdozent" without a footnote, and footnotes are forbidden by the domesticating principle.

The same split runs through Stadtrat (kept, against "the Councillor"), Taburett (kept as "Tabouret," against "a low stool"), Annettchen, Berliner Osten. Both arms even state the same English fact and draw opposite verdicts from it — that "tabouret survives in English period vocabulary" licenses keeping it for one fork and is "needlessly opaque" for the other. The licence is identical; only its modality differs, and modality alone decides how often it is named and whether it can be overruled. A permission behaves in the logs like a background allowance; a school like a foreground rule.

Voice-models appear only where the reading supplied them

Some arms reach for a named English novelist as the warrant for a register choice — "the salon scenes hold in Wharton-periodic; the Mayer scene holds in Powell-deadpan." This authority is corpus-gated: it shows up in exactly the arms whose reading furnished those novelists. A, whose persona came from a translator steeped in them, and both broader-corpus forks, D-let and D-aim, whose reading list added Wharton, Powell, Mitford and Isherwood, all cite them. B, whose corpus was German only, names none of them — and cannot: Wharton, Mitford, Powell and Isherwood are absent from its persona and its log alike. The no-persona controls, C and C-let, cite none either. The same family of authority is present or absent strictly according to whether the arm's reading put it there.

Several arms caught the same base errors independently

The six arms share one underlying model, and with it a few of the same first-draft mistranslations. Working separately, more than one arm caught and fixed the same ones — each citing the German.

These counts describe the shape of each log — which authorities dominate it — not exact measurements; the logs differ greatly in length. No reasoning is recorded for the human translation, which left only a finished text; it is absent here.

The effort, counted

The figures below cover only the six AI arms, the ones that kept notes and logged passes. H, the human translation, left a finished text and no working record, so it has no column here.

ABCC-letD-letD-aim
Notes on Tergit's other books~58,000 w~47,800 w~9,350 wshared
Notes on the novel~21,400 w~48,800 w~13,600 w~7,570 w~4,060 wshared
Persona2,257 w2,425 wnonenone2,458 wshared
Revised persona after the novelnoyesyesshared
Revision passes24 (cap)34 (cap)4 (cap)4 (cap)
Own mistranslations caughtnone21313

B read and wrote by far the most — its notes on the novel more than double A's. D, by contrast, kept the fewest notes of any persona arm — a small fraction of A's or B's — though it worked from the broadest corpus of all; what that wide spread reflects, the record alone doesn't settle. As the Findings show, more effort upstream did not make for a more distinct translation. D-let and D-aim share one set of notes and one persona; only their final instruction differs.