Findings
The analysis finds the human translation differs from the three machine translations sharply and consistently, while the machines differ comparatively little from one another. The gap runs across the chapter — in how names are kept or anglicized, idioms carried or remade, the dialect and the recurring refrain handled.
How far each pulls toward English
The three machines land within twelve points of one another, low on the scale — they keep the German close. The human sits sixty points off, at the far end.
The five measures behind the score
| Measure | A | B | C | H |
|---|---|---|---|---|
| Proper nouns | 0 | 40 | 60 | 80 |
| Loanwords & titles | 67 | 50 | 33 | 67 |
| Dialect | 0 | 0 | 30 | 100 |
| Refrain | 0 | 0 | 0 | 100 |
| Idioms | 25 | 25 | 25 | 100 |
| Composite | 18 | 23 | 30 | 89 |
Each value is the share naturalized, 0 (keeps the German) to 100; the composite is their mean.
The composite hides a split. The inherited arm keeps every German place-name (proper nouns, 0) yet naturalizes loanwords as freely as the human (67). On dialect, the refrain, and figures of speech the human goes the whole way over; the three machines barely move.
Pairwise agreement
Any two machines make the same call most of the time; a machine and the human rarely do — about three and a half times less often. The self-built arm and the control are the closest pair, agreeing on more than two-thirds of the loci. The human sits about the same distance from all three.
How each choice splits
On eighteen of the forty, the three machines agree and the human alone differs. The other twenty-two split every other way — but this set was built to include every cross-cutting case there is, so on an ordinary stretch the clean pattern runs higher.
Clustering
The self-built arm and the control are the closest pair; the inherited arm sits a little to their side; the human is far off. The persona each machine carried — inherited, self-built, or none — moved it only within the cluster, never across to her. The three also share one underlying model, the likeliest reason they group at all.
The inherited arm and the human
The inherited arm's persona was built from the human translator's own writing, so one might expect its English to lean toward hers. It does not. It agrees with her no more often than the control does; only two of the forty are uniquely hers — coupé → "carriage" and "Justice?", both taken from her essays — and on proper nouns it keeps the most German of all four.
The choices
The three machines mostly line up; the human mostly diverges. A few run odder: on Der Dämel! all four part ways — chump, great fool, ninny, sop; the control alone keeps the Berlin word Molle for a beer; the self-built arm alone capitalizes "You" to mirror the German Du of the love-letter; and the human alone rhymes the Schumann lyric (blind / mind) that the others give as prose.
| German | A | B | C | H |
|---|---|---|---|---|
| Süße … 1887 | What sweetness | What sweetness | What sweetness | How sweet the air was |
| Privatdozent | Privatdozent | Privatdozent | Privatdozent | lecturer |
| Einverstanden? | Agreed? | Agreed? | Agreed? | D'accord? |
| Frauenlieb und -leben | A Woman's Love and Life | A Woman's Love and Life | A Woman's Love and Life | Frauenliebe und Leben |
| Coupé | carriage | coupé | coupé | carriage |
| Das Recht? | Justice? | The law? | The law? | Justice? |
| Annettchen | Annettchen | dear little Annette | Annettchen | Annettchen |
| honor / theater | American | British | American | American |
| Schloss | Schloss | Schloss | the Palace | Berliner Schloss |
| Molle | beer | beer | Molle | beer |
| Graues Kloster | Graues Kloster | Grey Cloister | Grey Cloister | Gray Cloister |
| Feuerzauber | Magic Fire | Magic Fire Music | Magic Fire Music | Wagner's "Feuerzauber" |
| Kasinotoilette | dinner gown | casino gown | casino gown | casino toilette |
| Hammelbeine langziehen | box your ears | mutton-legs … haul | mutton-legs … stretch | good slap |
| wie ein Reh | fawn | doe | fawn | doe |
| Der Dämel! | The chump | great fool | The ninny | a sop |
The full dataset is in the Appendix.