A Second Chapter — Replication, and the Machines as Readers

Stance. The seven renderings are treated as co-equal columns. This is description and synthesis — not a ranking. Which rendering is the better or more faithful literary translation is a judgment reserved for human reviewers and is not made here. Every chart is built from the study's verified quantitative layer; every quoted rendering is from the finished translations.

1The design & the two probes

Chapter 26 of Effingers is a bourgeois Berlin Sunday — the extended Oppner–Effinger family gathering for the midday meal, the men herded off to their naps with a line from Lohengrin, the dinner-table talk threaded with operetta, embedded verse, period honorifics, and a quiet narratorial flash-forward to 1914. Its load-bearing features make translation choices visible: rank and kin honorifics (Justizrat, Herr, Tante/Onkel, Fräulein, Frau Kommerzienrat), untranslated culture-tags (tout Berlin, Mahlzeit, embedded magazine and operetta titles), Berlin idiom and the "die + surname" article, embedded lyric (Wagner, Goethe/Wolf, Schiller, the Zigeunerbaron pig-couplet), period epistolary register, and free indirect speech.

Six AI arms render the chapter, varying on two axes; the human translation is folded in as a seventh, co-equal column. The design is identical to the Chapter-25 study — the same arms, the same two clean probes — which is what makes this a replication rather than a fresh experiment.

Limitations, up front. n = 1 chapter, one instance per AI arm (a case study, not a statistically powered result). The six AI arms share a common base model — a plausible driver of their mutual closeness; the agreement figures measure how alike the outputs are, not independent convergence. This is an informed, not blind, comparison (arm identities known). And because the loci and their calibration are specific to this chapter, the absolute composites below are not comparable to the Chapter-25 numbers — only the structure is. All numbers are recomputed from the chapter's independently verified quantitative layer.

2The composite gradient — Chart 1

Scoring each rendering on a 0–100 foreignization↔domestication composite — here the discrimination-weighted mean of 41 scored loci, each locus weighted by how much it spreads the seven arms (its population standard deviation), so the honorific/title/culture-tag loci that do the separating carry the most weight — places the seven on a single line. H stands far out at the domesticating end; the six machine arms sit in a narrow band, and one of them, A, has drifted toward the middle.

3The 7×7 agreement heatmap — Chart 2

Two arms "agree" at a locus when they make the identical 0–100 pick. The matrix is symmetric; darker green = more agreement. The six machine arms form a dark-green block; the H row and column are a pale band — H sits apart from all of them, exactly as in Chapter 25.

4The cluster geometry — Chart 3

In Chapter 25 a second axis (lexical domestication × editorial boldness) opened the cluster into a 2-D scatter. Here that second axis does not survive the data: a classical multidimensional scaling of the distance matrix returns one dominant dimension and a faint, uninterpretable second. So the honest rendering is the line — the field is one-dimensional, and the one dimension is the foreignize → domesticate axis that Chapter 25's scatter largely recapitulated anyway.

5Region, not point — Chart 4

Is there a single "domesticated Tergit" target toward which stronger nudges converge? The two probes say no — and they say it through their vectors. A nudge moves an arm along the axis as a quantity; it does not aim the arm at H as a target. The permission's small step lands mostly toward H; the school's larger step splits — pulling some loci toward H and pushing as many away.

6Where the human sits — Chart 5

Domestication is not one dial. Comparing each arm's domesticating-mean to H's across six sub-axes (MORE / at / less than H) separates the dials the composite nets together. The single red cell tells the story — and this chapter relocates it.

H's structural and craft signature

What defines H is not more of what the machines do but a different kind of move. Across the 41 loci H is the sole most-domesticating arm at 23, and the "six machines agree, H alone differs" pattern recurs throughout. A cluster of moves belongs to H alone and to no machine arm — not even D-aim:

To these the layer adds the modernized period register throughout Paul's letter („häuslicher Herd gründen“ → "settle down," „äußerlich“ → "pretentious"), the stripping of Berlin's "die + surname" article entirely ("Mann," "Koller," "Widerklee" with no article, where even the most naturalizing machine keeps "the Widerklee"), interpretive paratext the German does not carry ("Hartert, the bank apprentice"; a footnote marker), and ready-made English idioms for German proverbs ("You are what you eat!").

And yet H foreignizes selectively — the anti-categorical signature, and the thing no machine reproduces. The same translator who Englishes the magazine title and rewrites the letter keeps the drink-word „Schoppen“ in German (L27) even while translating the pub-names around it, and renders „Brüsseler Poularde“ as the French menu-term "poularde de Bruxelles" (L24) — the single most foreignizing choice any arm makes anywhere in the chapter. H domesticates the references an English reader cannot catch and keeps the cultural artifacts an English reader is likelier to recognize, judged case by case. The machines are globally consistent — keep-German or gloss-it across the board; H is locally judged. That selectivity is itself unreachable by direction-nudge: D-aim, told to domesticate harder, landed on a different coherent profile — anglicizing the honorifics H keeps, keeping the realia H dissolves.

7How the arms reasoned — and read

This is the section the replication makes strongest, so it carries the weight of the page. The six AI arms left two kinds of artifact the human did not: reading notes written before any translation, and per-pass translation logs. The reading notes deliver the headline result; the logs explain why the finished outputs nonetheless diverge.

The machines read alike — a converged seminar

The hypothesis under test was that the machines might diverge in finished output yet converge in their reading. It is borne out, strongly. Reading the whole novel independently and writing private notes before translating, the six arms land on a near-identical inventory of Chapter 26's interpretive cruxes — not loose thematic overlap but the same sentences, flagged for the same reasons, often with the same instruction about how the translation must behave.

The convergence is sentence-level, not topical. The sharpest case is the Schiller doubled line („Denn mit des Geschickes Mächten ist kein ewger Bund zu flechten…“): all five reading-sets not only flag it but independently recover that the very same line is spoken by the bankrupt Mayer in Chapter 11, and all five issue the same cross-chapter rule — render both occurrences in identical English so the reader catches the echo. Five readers reconstructing the same buried novel-wide echo from a single chapter's vantage is like-minded reading at its most specific. The same holds for the bare "bis 1914" flash-forward, which every reader isolates and every reader prescribes rendering flat ("no italics, no flourish, no foreshadow-cue; the date does all the work"), and for the doomed Crown Prince, where every reader imposes the same self-denying rule that the irony "must not be said." The genuine divergences are minor and are the disagreements of close readers who agree on the text: Sofie's exact age, which beat is the pivot, how high to rank a particular pun.

Treated as readers, the six arms behave like a seminar that has converged — independently isolating the same flat date, the same recycled Schiller line two chapters apart, the same gesture-borne adultery, the same self-denying rule about the doomed prince. This is the AI as an excellent reader, and it is now the second chapter on which the result holds: beneath the divergent output, the reading is shared.

Where they diverge: the authority each arm is licensed to cite

If the reading is shared, the divergence in output is produced one layer up — in what each arm cites as its authority. An inductive typology of the pass logs sorts the warrants into a few kinds: a self (persona), the text, a permission, a doctrine, a fidelity-discipline. The cleanest signal is the persona-citation gradient:

It tracks exactly how much "self" each arm was given. A anchors in an authorial self ("per persona," "matches my own 'we are the last burghers'") and is the only arm to ground honorific policy in the persona. B anchors in a method persona ("the chorus-method," "the persona-note requires"). The controls C and C-let cite the persona zero or nearly zero times and reason from fidelity and the book's own recurrences — C-let adding the permission itself as a named warrant ("Domesticated, per style permission"). D-let reasons fidelity-first behind a stable of voice-models (Wharton / Isherwood / Mitford / Powell).

Probe 2 in the logs — the school changed the grounds, not the moves

With persona, reading, and instance held constant, what did the Schleiermacher instruction actually change in D-aim? Its stated authority and its vocabulary of justification — far more than its decisions. The school becomes D-aim's named top authority and displaces the persona-as-self (D-aim cites persona zero times; D-let cites it); it supplies a justification-vocabulary ("domesticating choice," "domesticated-equivalent," "domesticating-correct," "on the reader's own terms"); and it raises the reader to the deciding court (D-aim invokes "reader" 14× to D-let's 9×). On the Wagner line both arms reach the same standard English performance-form, but D-let justifies it by ear ("the slightly stiff English is the comic point") and D-aim by reader-doctrine ("it did not register to the English-reader as a Wagner-quotation"). Strikingly, the doctrine does not make D-aim strip the German: both keep Bürger in German, D-aim keeps more foreign markers and re-labels them as domesticating, and on every honorific tabulated D-let and D-aim are identical. The domesticating school changed the grounds offered, not the moves made.

The A anomaly — and the control that explains it

The new finding against Chapter 25 — A drifting to the most-domesticating machine — is not an emergent quirk of the translation pass. It is pre-written into A's persona. A's persona document, distilled from the human translator's own essayistic prose, states the rule verbatim: "honorifics simplified — the Prussian Rat titles all to 'Councilor,' 'Frau Kommerzienrat' to 'Mrs. Oppner'," inside a "mid-Atlantic" register that "naturalize[s] foreign society-words." A then cites that rule by name at the point of choice ("Honorifics simplified per persona… Justizrat → 'Councillor'"), and reads the whole novel from an inherited authorial stance ("my own work," "my family," "I am the author, not a faithful servant of the author"). Because this chapter is unusually honorific-heavy, that one pre-written rule is enough to carry A from the foreignizing end to A's drift toward H.

8Discussion

Two chapters now, one geometry. The replication confirms a structure and isolates a clean law beneath it.

The reading is shared — and that is the load-bearing result. On two chapters running, the six arms converge to ≈100% on the chapter's interpretive cruxes when they read independently before translating: the same flat date, the same Schiller line recovered two chapters apart, the same gesture-borne adultery, the same self-denying rule about the doomed prince. The divergence visible in the finished translations is not a divergence in comprehension. As readers the machines are a converged seminar; the differences enter only at the moment of rendering, and they enter as differences in licensed authority.

The geometry replicates. Six machines in a compressed band (27.5–52.7), the human alone at 69.3, an empty middle of 16.6 points to the nearest machine; machine–machine agreement 3.5× tighter than machine–H; every machine's nearest neighbour another machine; the field essentially one-dimensional along the foreignize → domesticate axis. Persona, permission, and school each move an arm within the cluster without crossing to H. This is the Chapter-25 result, re-run.

The probes behave as before — a nudge is a quantity, not a heading. Holding the persona constant, the school moved D-aim's composite by +9.2 and the permission moved C-let's by +2.8, both in the predicted direction. But the school's move splits 13/9 toward/away from H — it manufactures a generic "domesticated" surface rather than reconstructing H's specific editorial logic. Domestication-as-quantity is inducible; domestication-as-H's-specific-judgement is not.

The one new finding — and what makes it more than a curiosity. The arm that drifts toward H flipped from the school arm (Chapter 25) to the inherited-persona arm A (here), because the chapter changed: a honorific-heavy chapter activates the one rule A's persona pre-writes. The control makes the point general: the drift is not generic domestication (D-aim, told to domesticate, kept the titles) but the specific inherited human voice surfacing as a concrete lexical instruction and cited as authority at the point of choice. Give a machine a particular human's voice as its authority, and it moves toward that human — in its reasoning first, its output second.

The practical upshot is unchanged and now twice-observed: a categorical instruction can move an AI across the macro-regional axis but cannot specify the local picks within the region — the right pub name, the decision to fuse paragraphs, the rhyme that makes a misquoted lyric sound sung. The domesticating direction is a real and shiftable orientation; the destination is craft. What the human adds is not a dial-setting but a different repertoire — restructuring sentences, re-rhyming embedded verse, modernizing period register, substituting native idioms, pruning and keeping realia case by case — exercised on a reading the machines, remarkably, already share.

This is a description of an experimental space, not a verdict on it, and a deliberate replication of a prior chapter's design. Six of the renderings were made under controlled conditions for the purpose of comparison; one is a published human translation read in its own light. None of the seven is bad; each is a coherent reading of the German source. Limitations: n = 1 chapter, one instance per AI arm; the six AI arms share a common base model (a live confound for every similarity among them, and the most parsimonious read of the ≈100% reading-convergence is shared-base-model behaviour, not independent agreement among distinct readers); this is an informed, not blind, analysis; and the absolute composites are calibrated to this chapter and are not comparable to Chapter 25's — only the structure replicates. The chapter's comparative analysis (analysis.md) and the reasoning/reading exhibit (reasoning.md) carry the underlying detail.