Findings

Seven English versions of one chapter, measured against the German and against each other. The six machine versions cluster; the human stands sharply apart. And the new arms expose the deeper shape: "domesticated Tergit" is a region, not a point — push a machine harder toward natural English and it grows more domesticating without growing more like the human translator.

The measurements below come from a coding of 45 places in the chapter where the German leaves a real choice — names, idioms, dialect, titles, the spring-refrain, sentence shape. Two pairs of arms provide a useful control: C and C-let are identical but for a one-paragraph permission to domesticate, and D-let and D-aim are the same machine with the same persona, differing only in the instructions they were given at the moment of translation.

How far each pulls toward English

A composite of six measures, each weighted by how many of the chapter's coded choices it covers — the foreignizing–domesticating scale. 0 keeps the German; 100 naturalizes it into English.

The six machines occupy a narrow band, from 17 to 46. The human sits at 80, far out beyond every machine. The two most source-preserving arms are the ones carrying a persona on the lightest brief (D-let and B): left to itself, a persona pulls toward the German, not away from it. Only the school-instructed D-aim pulls clear of the machine pack — and even it stops well short of the human.

The six measures behind the score

Each value is the share naturalized into English, 0 (keeps the German) to 100; the composite is the six measures' coverage-weighted mean.

The composite hides the texture. The human goes the whole way over on dialect, structure, and idiom (100 on each) — the machines barely move there. But on lexicon the picture inverts: A naturalizes loanwords nearly as freely as the human (56), while keeping every German place-name (10). And D-aim is the only arm to climb on honorifics (83) — higher than the human herself.

The two probes — what the instruction alone does

Because two pairs of arms differ by a single input, the effect of the brief can be read directly.

A permission nudges the control gently toward the human. A categorical school instruction moves the dial far harder — the largest move in the study — and yet lands the translation no closer to the human's actual choices. This is the experiment's sharpest result: more domesticating is not the same as more like the human. A stronger instruction takes the machine to its own coherent destination, not to the human translator's.

Agreement: the field and the human

Any two machines make the same call about 63% of the time; a machine and the human, only 27% — a gap of more than two to one. Every machine-pair is closer than every machine-human pair: the tightest pair is B–C at 80%, and even the loosest machine pair (D-let–D-aim, 49%) out-agrees the closest machine-human pair (C-let–H, 38%). C-let is the machine that lands nearest the human; A sits mid-pack (27%) — no closer to the human translator than the others, though its persona was built from her own writing. (The six machines also share one underlying model, the likeliest reason they group at all.)

How each choice splits

On 14 of the 45, the six machines agree and the human alone goes another way — the largest single pattern, and the spine of the machine-versus-human divide. But most of the chapter, 31 loci, splits other ways: the new arms cut the field into shifting groups rather than one clean machine bloc, especially on titles, honorifics, and idioms, where the permission and the school pull different arms apart.

Clustering

The six machines sit together in the lower-left: they domesticate little and reshape form barely at all. The human is alone in the upper-right. The two axes come apart for the machines: D-aim moves right (it domesticates vocabulary and titles) but hardly rises (it leaves the German's form intact). Editorial boldness — fusing the refrain, inventing a rhyme, writing dialect — is, in this chapter, something only the human does.

Region, not point

When the domesticating arms abandon a vivid German idiom, they do not converge on one English answer. They scatter — each picks a different destination.

At the loci where all four domesticating arms leave the German behind, they pick on average 3.3 different English destinations out of a possible 4 — nearly maximal scatter. They share the direction — leave the German image behind — but split on the destination. There is no single "domesticated Tergit" the machines converge toward; the human translator is one path through a wide space of valid English, and an instruction moves a machine into that space without moving it to her particular point. (The one place they converge — "cow" for doofe Ziege — is the honest exception: a flat insult has an obvious English match; a vivid idiom does not.)

Where the strongest instruction lands, relative to the human

Domestication is not a single dial. Broken into its sub-axes, every machine arm sits less domesticating than the human almost everywhere — with exactly one exception in the whole grid.

Each cell: more / at / less domesticating than the human (H) on that sub-axis.

The lone MORE in the table is D-aim on honorifics — it alone turns Fräulein and Herr into Miss and Mr, a move the human herself never makes. Everywhere else the human is at or beyond the machines' reach — at the ceiling on dialect, structure, and locale, where no machine follows. The strongest-instructed arm overshoots her on one narrow dial while falling short on every other: it becomes not the human translator, but differently domesticating.

What only the human does

Five moves recur in the human translation that no machine arm makes — not even the most strongly instructed:

• Folds the spring-refrain into the surrounding prose; every machine keeps it standing alone as a beat.
• Recasts the refrain's "What sweetness" into "How sweet the air was" — inventing a subject and a verb the German doesn't give.
• Finds an English rhyme for the Schumann lyric (blind / mind) where the German, as the girl misquotes it, has none.
• Writes Berlin street-dialect in eye-dialect ("d'you haff to," "leggo," "yer old man"); every machine refuses it as a matter of stated principle.
• Breaks a character's free-indirect thought into the first person — "to be honest about my desires" — where the machines keep the German's impersonal frame.

And yet the human is no across-the-board domesticator. She keeps Fräulein and Herr; she leaves casino toilette in French — the most literal rendering of that word among all seven. Her domestication is decided case by case, on the ground of what each moment needs — which is what a categorical instruction cannot supply, and why the school-instructed machine moves in her direction without arriving where she is.

The choices

A sample of the loci, with each version's rendering side by side. The machines mostly line up; the human mostly diverges — and where the domesticating arms move, they move apart.

The full 45-locus dataset, the agreement matrix, and the measure-by-measure scoring are in the Appendix, alongside the complete one-page study with every chart.