1 Comparative Voice Engineering as Editorial Method

First Pair, LitVoice, and the Rosetta format in Lighthouse Republics and the Russian edition of The Invented Enemy

First Pair Editorial Research
Methods article · 31 July 2026

1.1 Abstract

This article describes a reproducible editorial method developed across two book projects. Lighthouse Republics began with a large civic-documentary manuscript whose “South” voice drew selected rhetorical mechanisms from Joaquim Francisco de Assis Brasil’s 1882 history of the Riograndense Republic: the returned witness, geographic force, periodic syntax, material scene-setting, and direct judgment of public institutions. Two sibling transformations were then made from the common Full Tide 2.0 text. The “East” experiment used short declaratives, ordinary physical action, understatement, and indirect reflection; the “West” experiment used spare coordination, concrete objects, submerged explanation, and dry treatment of loss. Production records originally named Haruki Murakami and Ernest Hemingway. Current practice retains those names only as archival shorthand and describes the transformations through neutral, high-level craft properties.

The experiments led to LitVoice, a rights-aware, provider-neutral toolkit that separates source discovery, rights decisions, private aggregate analysis, public-corpus N-grams, rewrite planning, originality checks, and human authorship. They also led to Rosetta: a format for two or more aligned editions with compact rails for divergence, feature vectors, structural marks, and exact equality; a three-pane Rosetta is a triptych. Rosettas are delivered as reflowable Kindle EPUBs and as bespoke Obsidian vaults. Desktop vaults maximize simultaneous comparison and editorial decision-making; mobile vaults preserve the same aligned evidence through horizontal, snap-scrolling panes. Structured controls can record decisions at chunk, sentence, word, and phrase scales without changing any source edition. The Enemy profile can compose an R-first proposed revision, while the Venice profile emits an instruction digest for a separate next-edition pass.

The second case study applies the method to three Russian translations of The Invented Enemy: Fable (F), Codex Review (R), and Codex (C). The review combined sentence-level source checking, a rights-scoped Russian literary corpus, word and character N-grams, function-word and part-of-speech patterns, rhythm and syntax diagnostics, a segregated translated-Russian simplicity device, and claim-level fact verification. The final instrument contains 2,633 losslessly aligned chunks across 25 chapters. Its scores are deliberately plural: Russian surface obstacles, inter-edition divergence, corpus distance, and human evidentiary judgment remain separate. The central methodological claim is simple: measurement is most useful when it directs attention without pretending to replace reading.

Keywords: digital scholarly editing; version comparison; literary style; translation revision; N-grams; Russian prose; Obsidian; EPUB; Kindle; reproducibility

1.2 1. The edition as an experimental object

Most editorial workflows collapse comparison too early. A writer, translator, or editor keeps one preferred file open, consults an earlier file only when doubt arises, and eventually loses a reliable account of what changed, where it changed, and why. Version control solves the archival problem but not the perceptual one: a Git diff can show inserted and deleted strings, yet it cannot by itself show three coherent voices as reading experiences.

First Pair treats an edition as both a publication state and an experimental object. Each named state is pinned to a source inventory and a commit; each comparison unit is aligned across states; each diagnostic is attached to the text it describes; and each durable decision returns to the canonical manuscript. The comparison surface is regenerated, not hand-maintained. This creates a loop:

1. freeze source states;
2. align their conceptual or textual units;
3. compute limited, interpretable diagnostics;
4. read the alternatives together;
5. make a human editorial decision;
6. edit the source and rebuild every derivative.

The loop matters more than any individual score. Its purpose is to preserve the connection between judgment and evidence.

1.3 2. Lighthouse Republics: from the South voice to three families

1.3.1 2.1 The civic-documentary foundation

The voice now called South emerged during the Full Tide expansion of Lighthouse Republics. The historical source charter names Assis Brasil’s Historia da Republica Riograndense (1882) as a principal craft debt, joined to the traditions of Michelet and Bancroft. The method did not copy Assis Brasil’s sentences or adopt his politics. It extracted a limited editorial repertoire:

• a narrator who returns from the archive as a witness;
• geography treated as a political force rather than scenery;
• material details—rope, ink, timber, mud, water, cannon, paper—used to ground abstraction;
• long periodic sentences capable of holding evidence, qualification, and verdict together;
• direct civic judgment, with institutions treated as moral actors;
• hindsight used as declared structure rather than concealed omniscience.

The modern South charter also states the negative boundary: no copied sentences, racial hierarchy, geographic determinism, imperial nostalgia, or claim that one archive exhausts the subject.¹ “Assis Brasil-derived” therefore means a documented selection of mechanisms, not simulation of an author.

The sequence is recoverable from commits and release ledgers. A 5 July 2026 expansion established the Rio Grande-derived historical energy; the 6 July Full Tide 2.0 release made the approximately 206,000-word, ten-part manuscript and its voice charter explicit. Later South revisions altered sentence density, narrator self-announcement, chronology, source apparatus, and scale. The Fourth Edition, committed on 13 July, restored the complete South body while tightening paragraph architecture and preserving the book’s documentary amplitude.²

1.3.2 2.2 “Murakami” and “Hemingway” as archival names

On 6 July, two complete transformations were produced from the Full Tide common source. The production branches and early charters called them “Murakami” and “Hemingway.” Those names truthfully describe the workshop prompts of that moment, but they are not suitable claims of equivalence, authorship, or successful imitation.

The stable family names are East and West:

East is not “what Murakami would have written,” and West is not “what Hemingway would have written.” English-language examples also mix author, translator, and editor signals; they cannot establish Japanese syntax or isolate individual authorship. Current project policy therefore treats the old names as historical version strings and publishes only the neutral craft descriptions.³

There is a second caution. South, East, and West are siblings in a common-source tree, not a clean sequence in which one variable changed at a time. Subsequent branch work introduced factual corrections, structural changes, chronology, illustrations, and source apparatus at different moments. A measured difference among the families may therefore be a voice difference, a later correction, or both. The vault records ancestry precisely so that the reader does not mistake a historical branch comparison for a controlled experiment.

1.4 3. Why the experiments became LitVoice

The first charters were useful but insufficiently formal. They mixed aesthetic observations, production instructions, source rights, narrator rules, and generation prompts. LitVoice was created in July 2026 to separate those concerns.

The project’s first invariant is rights. Bibliographic discovery may be broad, but source bodies enter a retained model only when their rights basis is explicit. A text being readable online does not authorize downloading, modeling, redistribution, or imitation. The software therefore distinguishes:

• metadata discovery;
• rights-cleared public or open-license acquisition;
• privately authorized, local-only analysis;
• retained lexical models;
• non-retained aggregate calibration;
• rewrite planning;
• the external act of authorship.

The second invariant is provider neutrality. LitVoice does not contain credentials, a model SDK, a network generation client, or a default generator. It prepares bounded requests from caller-supplied prose, preserves source hashes and chunk boundaries, audits narrator authority and structural fidelity, and evaluates returned candidates. A human, Fable, Codex, a local model, or another authorized host can supply the prose.⁴

The third invariant is originality. The toolkit’s neutral profiles—Restrained Uncanny and Anchored Interiority—describe abstract craft behavior rather than an author identity. Anchored Interiority V3, for example, plans first-person attention through ordinary action, perception, association, uncertainty, and return to the physical scene. A fixed candidate rubric weights fidelity, anchored interiority, restraint, viewpoint, cadence, and originality; fidelity, viewpoint, and originality have hard gates. Rhythm calibration can promote only stable, coarse controls. Private inputs leave no excerpts, vocabulary, N-grams, source identifiers, or retrieval index in the packaged calibration.

Public-domain and explicitly open-license corpora use a separate retained N-gram path. It supports deterministic word models through order five, corpus fingerprints, rights manifests, build reports, and copy-overlap guards. That lexical path is intentionally unavailable to private in-copyright collections.

This architecture grew directly from the Lighthouse work: an informal voice charter became a typed profile; a large rewrite became a source-authoritative batch plan; an intuitive impression of rhythm became a bounded aggregate envelope; and three unwieldy manuscripts became a comparative editorial reader.

1.5 4. The Rosetta format

1.5.1 4.1 Alignment before visualization

“Rosetta” names the comparison instrument, not a model. Its fundamental operation is alignment.

The normative cross-project specification is the First Pair Rosetta guide. It defines Rosetta as a two-or-more-pane format and reserves triptych for a three-pane comparison. Each build binds ordered source identities, an alignment contract, pane order, control and decision schemas, artifact hashes, and stable review targets. Generated evidence and human decisions remain separate: the reader may be rebuilt, while its structured ledger is preserved or rejected explicitly when source identity changes.

In the Lighthouse research vault, a source record is a sense unit: a paragraph, quotation, list, table, code block, or figure in one edition. A triptych fragment or slice represents one conceptual lineage and may concatenate zero or more sense units in each pane. This lets a concept remain comparable when one edition splits a paragraph and another fuses it.

The Russian translation instrument uses a stricter aligned sense chunk. Corresponding headings align by stable order inside each chapter. Prose blocks then align within heading boundaries through bounded dynamic programming over Russian content-word overlap, light stemming, sequence, block kind, and length. Aligned spans are refined into sentences and clauses, then packed only while every F/R/C slice remains within the configured Kindle budget. The builder reconstructs every source stream from the resulting chunks and refuses the build if any heading or prose block is lost, duplicated, or reordered.⁵

1.5.2 4.2 Lighthouse scores: divergence, not merit

Each Lighthouse rail begins with a single edition letter—S, E, or W—followed by a 0–100 divergence score and a seven-position feature vector. The score combines:

• 55% text dissimilarity;
• 15% length deviation;
• 15% sentence-count deviation;
• 10% unique-token pressure;
• 5% unique rare-token pressure.

All-equal pane groups are forced to zero. A larger number means that the pane is more different from its neighbors; it does not mean that it is worse. The = feature can also mean merely that no N/M/S/R/E/C/F threshold fired, so only the explicit equality banner or group proves normalized textual identity.

The vector occupies fixed positions so that a blank is informative:

The mark cell supplies another scale: no bar for shared wording, | for local pressure, || for substantive addition, omission, or a new claim, and ||| for a review outlier. A dotted rule marks structural split, fusion, or reorder.

The first page makes the value of simultaneous reading obvious. South opens with one long civic period; East breaks the scene into ordinary declaratives; West compresses it into a spare parable. A diff would emphasize character changes. The triptych exposes three reading experiences while the rails point to likely mechanisms.

1.5.3 4.3 Equality as editorial information

Exact equality is independent of scores and feature vectors. In the Lighthouse desktop reader, equality is recalculated among panes currently visible. In the Russian instrument, equality collapses whitespace but preserves case and ё; it appears as F=R, F=C, R=C, or F=R=C.

The vertical chain is intentionally redundant. A compact rail that reads

F
=
R
=
C

lets the reader skip three equal slices without scanning their prose, while a pair relation such as F=C immediately isolates the independent revision.

1.6 5. From comparison document to editorial environment

1.6.1 5.1 The full desktop vault

The full Lighthouse Republics vault is a research graph, not merely a reader. Its current build contains 29 immutable releases, 50,819 sense units, 49,012 direct transformation events, and 2,947 conceptual lineages in 117,432 files. Edition notes link to people, places, dates, sources, figures, transformations, and lineages. A user can read one release, inspect one parent-to-child change, or follow a concept through the whole release tree.

The desktop triptych plugin layers reading controls and editorial decisions:

1. visibility chips show or hide an edition globally; visibility is a reading setting, not a decision;
2. prefer this is mutually exclusive across the panes of one fragment and records the whole-fragment choice;
3. an editing note records a compact imperative for the next edition;
4. hidden-by-default sentence choices record independent sentence preferences, rather than forcing false one-to-one exclusivity across split or fused prose;
5. a phrase selected inside one sentence can be marked Prefer this or Ungrammatical;
6. a persistent decision ledger can be rendered as a next-edition digest.

The structured state lives in Preferences/triptych-decisions.json; Preferences/EDITING-INSTRUCTIONS.md is its portable human-readable digest. Venice records evidence and instructions rather than composing a manuscript inside Reader. The ledger is deliberately separate from generated comparison notes, so the complete desktop and mobile vaults can refresh source evidence without erasing human decisions.

1.6.2 5.2 The preview vault

The full graph is too large for casual distribution. The Preview Vault is a curated, self-contained research edition containing the complete front matter and Part I across four strategically chosen states: Full Tide (Fable), South 2.2.1 Final, Third Edition Revised, and Fourth Edition. Its current scope is 743 units, 119 transformation events, 276 lineages or triptych fragments, and 1,692 files—large enough to demonstrate the method, small enough to inspect as a coherent edition. It is a clean replacement build: regeneration replaces its Preferences/ directory, so it is not the sole durable store for private editorial decisions.

The selected fragment is methodologically useful. Full Tide, South, and the Fourth Edition are textually identical; the Third Edition is a substantial skeptical rewrite. Its score of 66 and three-bar outlier mark do not reject it. They state that a decision is required. The human question is whether the explicit limit—“Mud, however, enacts no political law”—is a valuable correction to an overextended metaphor or an interruption of the book’s governing voice.

1.6.3 5.3 Mobile is a different edition, not a shrunken desktop

The full vault is inappropriate for a phone. The Lighthouse Mobile Vault contains all 29 named editions and all four configured triptychs but omits the archival transformation graph. Its current manifest records 3,988 edition pages, 50,819 units, 8,609 triptych slices, 214 compact source notes, and 905 files. Edition prose is stored once and loaded on demand.

Mobile comparison changes the spatial grammar:

• vertical motion advances through slices;
• horizontal motion changes edition pane;
• tabs jump directly to a pane and follow a swipe;
• each pane leaves a visible sliver of the next;
• static Previous, Chapter, Top, Contents, and Next links preserve navigation without the plugin.

Mobile shares the sentence- and phrase-level review controls. Whole-fragment preference and editing-note entry remain desktop controls in the present Venice profile. The desktop surface answers, “What can I compare at once?” The mobile surface answers, “How can I keep my place while reading one alternative closely?” Neither is a degraded version of the other.

1.7 6. Translating and reviewing The Invented Enemy

1.7.1 6.1 Three Russian manuscripts

The Russian edition created a second use for the Rosetta system. Three complete translations shared the same English source and chapter architecture:

The review edition did not seek a compromise sentence halfway between F and C. It returned to the English source and wrote Russian sentences under an explicit editorial charter: preserve argument, evidence, dates, uncertainty, and rhetorical force; remove English information order, literal connective tissue, excess nominalization, and bureaucratic caution; allow Russian word order, sentence division, and cadence to reorganize the surface.

An early charter named Dostoevsky as an orientation for moral pressure and argumentative self-examination. As with the Lighthouse experiments, the publishable claim is narrower: the review sought energetic intellectual prose, direct questions and reversals, and a sentence ending with a clear stress. It did not seek nineteenth-century vocabulary or a fictional Dostoevsky pastiche.

1.7.2 6.2 Review as a sequence of different tests

The review separated tasks that are often conflated:

1. semantic comparison against the English source;
2. Russian copy-editing for syntax, collocation, rhythm, repetition, and register;
3. structural invariants for headings, links, emphasis, tables, codes, dates, and numerals;
4. corpus diagnostics against rights-reviewed Russian prose;
5. translated-Russian simplicity against a private aggregate device;
6. claim verification for every live [verify…] marker and Appendix A chronology flag;
7. artifact validation for PDF, EPUB, Kindle Rosetta, and Obsidian vault.

This order matters. A fluent mistranslation can score well. A faithful sentence can be needlessly foreign. A historically elegant sentence can still make a false claim. No single model can adjudicate all three.

The verification pass turned each marker into one of four outcomes: verified, corrected, qualified, or removed. Eleven batch claim ledgers, governed by a verification README, preserve the original claim, primary or authoritative evidence, counterevidence, locator, and disposition. The book chapters and bibliography now contain no live verification tokens or draft banners. Historical research dossiers still mention the old marker convention because they preserve the project’s research history; those archival mentions are not unresolved manuscript instructions.⁶

1.8 7. A Russian LitVoice corpus with explicit boundaries

1.8.1 7.1 The public literary handoff

The Russian LitVoice project lives outside the book repository. It owns the rights manifest, acquired texts, models, style devices, and benchmark reports; book projects consume the reports without vendoring the corpus.

The positive-weight prose handoff contains fourteen works by ten authors: Tolstoy, Dostoevsky, Bunin, Chekhov, Turgenev, Pushkin, Lermontov, Gogol, Goncharov, and Leskov. Authors receive equal aggregate weight, and works are length-weighted within each author. A Yesenin collection is held as a separate poetry register rather than mixed into the prose envelope.

The public-corpus build produces:

• word 3- and 5-grams;
• character 3-, 4-, and 5-grams;
• function-word 2-, 3-, and 4-grams;
• morphological part-of-speech 2-, 3-, and 4-grams;
• sentence, clause, punctuation, rhythm, nominalization, passive, participial, gerundive, and verbal-energy profiles;
• chapter- and sentence-level outlier reports.

The current handoffs contain 710,947 positive-weight prose words and 17,451 separately registered Yesenin words, for 728,398 Russian words in all. Every build is bound to the source manifest, byte-stable acquisition ledger, page-revision lock, content audit, normalized streams, dependency lock, LitVoice checkout, and implementation fingerprint.⁷

1.8.2 7.2 Copyright and artifact availability

“Russian classic” is not itself a rights category. The project makes a United States-scoped determination per work and separately requires a lawful, reproducible artifact.

This produces results that may look conservative:

• the exact 1917 first edition of Pasternak’s Поверх барьеров is acquired from a pinned scan and OCR derivative but has model weight zero because historical OCR should not shape the authenticity envelope;
• other Pasternak versions remain excluded until textual version and rights are separately verified;
• Paustovsky’s eligible early works remain excluded because no lawful, reproducibly downloadable artifact was pinned;
• Grossman remains outside the public corpus under the project’s conservative U.S. screen, even though a Soviet or Russian copyright calculation may differ.

This is not a claim about global copyright and not legal advice. It is a reproducibility rule for this model. A privately authorized Grossman omnibus can enter the segregated local analysis lane without being redistributed or represented as public domain.

1.8.3 7.3 Private books and the simplicity device

Privately held FB2 and EPUB books are read in place. The extractor excludes metadata, headings, notes, binaries, covers, navigation, and encrypted content. It persists no prose, passage, vocabulary, or retrieval index. A local provenance record retains pseudonymous source IDs and content hashes for reproducibility while stripping paths, titles, and authors.

One derived device, murakami_simple_ru, describes the simplicity of a declared set of Russian translations through four non-lexical dimensions:

• short syntactic arcs;
• a concrete-action diction proxy;
• restrained modifiers;
• paragraph motion.

The reusable device retains only group-balanced aggregates, with no source identities or hashes. Its fit index means proximity to that translated-Russian simplicity envelope. It is not native-Russian authenticity, a translator classifier, a Murakami imitation target, or a publication verdict. This separation lets an editor ask for “more simplicity” without treating copyrighted translations as a reusable lexical voice.

1.9 8. The Russian Rosetta

1.9.1 8.1 Two scores that must not be merged

The F/R/C rail adds a second score family.

РУ is a deterministic surface-obstacle heuristic. It begins at 100 and subtracts penalties for configured patterns: literal calques, bureaucratic or nominalized prose, overloaded sentences, intrusive repetition, monotonous rhythm, syntactic heaviness, excessive Latin-script density, and chains of clipped fragments. A high score means that the detector found fewer of those obstacles. It is not an objective fluency or literary-quality score.

Δ independently measures divergence from the other translations: 82% text dissimilarity and 18% length deviation. A pane can therefore have РУ 100 and still be a Δ outlier. That combination is often editorially interesting: a fluent sentence may have changed the claim.

The Russian feature vector uses:

The Kindle build uses one table-native comparison per EPUB spine item. Edition letter, РУ, Δ, feature vector, bars, structural rule, and equality chain remain physically attached to their text under reflow. The current validator checked all 2,633 items and calculated a maximum of 26 Scribe lines per slice.

1.9.2 8.2 Current alignment and equality

The clean current build is version 1.1.0-957c4b6-9377f44-5b4bf57: Fable at 957c4b6, Codex Review at 9377f44, and Codex at 5b4bf57. It contains 25 chapters and 2,633 aligned chunks.

The aggregate triptych report gives F РУ 95, mean Δ37.34, and 10 outliers; R РУ 98, mean Δ48.08, and 520 outliers; C РУ 95, mean Δ39.80, and 36 outliers. The defensible reading is not “R wins.” It is that R triggers fewer simple surface detectors while being much more independently recast.

1.9.3 8.3 Desktop and mobile translation vaults

The complete desktop and mobile vaults contain the full 2,633-chunk comparison and the same provenance material. Their pinned Reader navbar provides Previous, Up, Back, Top, Contents, and Next; generated Markdown links preserve navigation when the plugin is disabled.

The desktop layout stacks F, R, and C vertically. A narrow left rail remains visible beside each full-width sentence, including a vertical equality chain. This gives long Russian prose more measure than three narrow columns would.

The mobile layout uses horizontal snap scrolling. The rail moves to the top of the card; F/R/C tabs follow taps and swipes; the next pane remains slightly visible.

The passage demonstrates why divergence is not a penalty. F and C say that a ballad prepared or inflamed the war Sweden later lost. R distinguishes what the evidence supports: the killing entered the official justification for war, but the ballad’s decisive causal role is a durable research legend rather than an established fact. Δ82 correctly calls for attention; historical review determines that the divergence is necessary.

The private editable preview is a separate nine-page laboratory covering canonical chunks 27–35. It exercises the full Reader/Review workflow while keeping complete-book delivery stable. Each page exposes an always-visible, mutually exclusive Overall F | R | C choice. Sentence & words reveals sentence preferences and one Words button per nonempty pane unit. Inside that surface, Whole sentence selects the unit, adjacent word checkboxes select a contiguous phrase, and Only selects a single word.

A selected phrase can be marked Good, Bad, or Edit. Saved phrases may overlap, nest, or cross-overlap; each remains independently addressable. An edited highlight reopens its variants beneath the action row. Save & replace changes the preferred wording in place; Save next appends another timestamped variant; stacked checkboxes select exactly one preferred variant for composition. An actual heading page carries a Heading badge, allowing a title such as Слово с двумя назначениями to be edited as the structural heading rather than as repeated navigation context. Decisions are serialized separately from the generated pages, with a JSON ledger, integrity sidecar, and human-readable digest. There is no ordinary “Preserve previous choices” control: source-mismatch recovery appears only when nonempty saved decisions genuinely need archival treatment.

1.9.4 8.4 Review and R-first assembly

Review is a peer view beside Reader. It lists only pages with saved decisions, links each card back to its source page, and shows a complete New copy preview for the vault’s scope. The normal base is R, so an editor changing only a few phrases does not need to prefer R repeatedly. Composition precedence is sentence preference, then whole-chunk preference, then the default R copy; explicitly mapped phrase replacements are applied last.

Phrase insertion is deliberately explicit. Bad is feedback only and can never enter the proposed copy. A Good or Edit annotation becomes operative only after Use in copy maps it to an exact destination in the currently effective base text. Good inserts the selected source wording; Edit inserts the checked preferred variant. Destination mappings may not overlap, and stale or ambiguous targets fail closed. This makes a typical revision legible: retain R globally, cherry-pick exact F or C phrases where they are better, revise an R phrase directly when necessary, and inspect the assembled result before export.

Save review copy writes Markdown plus JSON provenance without modifying F, R, or C. In a complete vault it can yield a full proposed next R revision; the current editable preview exports only its bounded excerpt. Promotion into canonical R remains a separate reviewed editorial act, followed by rebuilding the book and all Rosetta surfaces under a new source identity.

1.10 9. Iterative measurement and what changed

The pre-edit Codex Review snapshot (R0, commit a98afa5) and the verified current review (R, commit 4bf0e00) are unusually useful controls because they share genre, source, and architecture.

The outward movement in classics distance is not a failed revision. The final verification pass added names, titles, dates, citations, qualifications, and modern analytical prose—the very materials that differ from a nineteenth-century fiction envelope. Direct translationese fell sharply, while function-word scaffolding and rhythm remained substantially closer to the corpus than the other two complete translations.

The current three-way benchmark is:

Each column answers a different question. Classics distance is weighted divergence outside an author-balanced historical-literary envelope. Its display index is merely 100 / (1 + distance / 25). Translationese counts a narrow pattern list. The simplicity device measures proximity to a private translated-Russian aggregate. None tests factual truth, semantic fidelity, native authorship, or publishability.

The iterative method therefore uses quantitative results comparatively:

• compare like-for-like drafts;
• inspect feature families, not only one total;
• read high-scoring and low-scoring sentences;
• check every proposed edit against the source;
• preserve counterexamples and deliberate departures;
• rerun after factual revision, because changed content changes style statistics.

1.11 10. Discussion

1.11.1 10.1 A score should create a question

The strongest design decision in the Rosetta system is that no score is allowed to stand alone.

• A Lighthouse divergence number is accompanied by its panes, feature positions, bars, and structural mark.
• A Russian РУ score is accompanied by Δ, equality, and a diagnostic vector.
• A classics-distance total decomposes into word, character, function-word, POS, rhythm, and verbal-energy families.
• A simplicity fit is explicitly segregated from native authenticity.
• A factual change is documented in a claim ledger.

This prevents a common category error: using a measurable proxy as the name of an unmeasured value. “Different” is not “bad.” “Closer to historical prose” is not “more Russian.” “Simple like this translation aggregate” is not “authentic.” “Fluent” is not “faithful.” The system’s purpose is to put the right question beside the sentence.

1.11.2 10.2 The reader is part of the method

Format is not an afterthought. The same evidence behaves differently on different devices.

The Kindle edition gives each conceptual comparison a pagination boundary and keeps the rail attached under reflow. Desktop vaults give alternatives a common visual field; mobile vaults trade simultaneity for stable place, direct tabs, and close reading. Preview has a profile-specific meaning: Venice offers a bounded clean scholarly companion, while Enemy uses a private nine-page control laboratory. The decision layer is likewise profile-specific. Venice compiles its ledger into an instruction digest for a separate edition pass; Enemy Review composes an inspectable R-first candidate from explicit chunk, sentence, and phrase choices.

These are editions in the full editorial sense: selections and arrangements designed for different acts of attention. In both profiles, generated reading evidence, persistent decisions, and canonical source prose remain distinct.

1.11.3 10.3 Reproducibility includes refusal

The projects fail closed in several places:

• a source on the wrong branch;
• a dirty chapter inventory without an explicit snapshot mode;
• different heading architecture;
• lossy reconstruction;
• an overlong Scribe slice;
• stale provenance between EPUB and vault;
• an unknown or insufficient rights basis;
• a private corpus leaking expressive data;
• a score presented as a stronger claim than its definition permits.

Refusal is not incidental friction. It is part of the research method because it keeps the comparison tied to identifiable sources.

1.11.4 10.4 Limitations

The method has clear limitations.

First, the Lighthouse branches were not controlled experiments. Voice, fact, structure, and apparatus sometimes drift together. Second, alignment is interpretive even when deterministic; a conceptual lineage can combine or omit source units. Third, all fluency and authenticity diagnostics are partial and domain-sensitive. Fourth, English translations of Japanese prose cannot isolate an author’s original syntax. Fifth, the private simplicity device has uneven translation groups and is descriptive rather than cross-translator calibration. Sixth, the Russian literary corpus is historical fiction used to interrogate modern analytical nonfiction; absolute distances are necessarily dominated by genre and period.

Finally, the application figures in this article are controlled renderer captures. The EPUB images use production package XHTML and CSS at a Scribe-scale viewport. The vault images use production payloads, DOM class structures, and plugin CSS inside a stable theme shim. They document the actual artifact logic, but they are not photographs of Kindle firmware or literal macOS screenshots.

1.12 11. Conclusion

First Pair’s editorial method arose from a practical problem: how to keep several complete, plausible books visible long enough to learn from their differences.

Lighthouse Republics supplied the first case. A South civic-documentary voice was made explicit, two sibling transformations exposed different sentence-level pressures, and the need for rights-aware abstraction led to LitVoice. Rosetta then turned branch history into an instrument: aligned prose, divergence rails, feature vectors, equality relations, and human decisions in one field.

The Invented Enemy supplied the harder second case. Translation review required not only style comparison but source fidelity, Russian idiom, corpus evidence, private-data separation, and claim verification. Its F/R/C Rosetta shows why the system keeps axes distinct. The reviewed translation can be more fluent by one heuristic, more divergent by another, closer to a historical corpus on some feature families, farther after factual expansion, and still require sentence-by-sentence human judgment.

The broader lesson is not that literary editing has been automated. It is that editorial attention can be engineered. A good instrument remembers provenance, exposes alternatives, decomposes its measurements, and adapts to the reader’s device. It also makes post-processing explicit: Venice produces a decision digest for the next-edition editor; Enemy produces a provenance-bearing composed-copy candidate based on R and selected alternatives. Neither silently edits a source edition. Only a separate reviewed promotion, followed by a complete rebuild, returns the accepted decisions to canonical prose.

1.13 Repository and reproducibility notes

The article reports local project states as of 31 July 2026. Primary materials are:

1. Lighthouse Republics release and prompt ledgers: RELEASES.md, PROMPTS.md, scripts/obsidian_snapshot.py, and docs/VENICE-VAULT.md in ~/src/venezia/usavenice.
2. Historical voice charters: notes/research/voice_charter.md in usavenice-codex-brasil, murakami_charter.md in usavenice-codex-murakami, and hemingway_charter.md in usavenice-codex-hemingway.
3. LitVoice architecture and rights policy: README.md, docs/RIGHTS_AND_RESEARCH.md, docs/REWRITE_ENGINE.md, docs/V2_ENGINE.md, docs/V3_RHYTHM_ENGINE.md, and docs/NGRAM_VOICES.md in ~/src/litvoice.
4. Lighthouse readers: scripts/build-obsidian-triptychs.py, obsidian-plugin/lighthouse-triptych/, scripts/build-obsidian-mobile-vault.py, obsidian-plugin/lighthouse-mobile-reader/, and docs/VENICE-MOBILE-VAULT.md.
5. Russian corpus and reports: README.md, corpus/sources.json, reports/russian_corpus.audit.json, and generated benchmark artifacts in ~/src/russian.
6. Translation methods: Rosetta.md, docs/TRANSLATION-RU.md, docs/TRANSLATION-TRIPTYCH.md, docs/TRANSLATION-TRIPTYCH-EDITABLE-PREVIEW.md, scripts/build_translation_triptych.py, scripts/build_translation_triptych_vaults.py, scripts/build_translation_triptych_review_preview.py, and sources/verification/ in the present repository.
7. Current translation triptych report: version 1.1.0-957c4b6-9377f44-5b4bf57, built and validated from three clean commits on 31 July 2026.
8. Canonical cross-project format specification: First Pair Rosetta guide, commit 6cef770.