> ## Content Index
> Fetch the complete content index at: https://firstpair.press/llms.txt
> Use this file to discover other available public pages before exploring further.

# Rosetta: from comparative reader to editorial instrument
- URL: https://firstpair.press/comparative-voice-engineering-firstpair/
- Published: 2026-08-01T01:09:24.000Z
- Updated: 2026-08-01T01:09:24.000Z
- Description: Rosetta now joins aligned editions to source-bound chunk, sentence, word, phrase, heading, variant, and Review workflows for editorial revision.
- Author: Alexy Khrabrov
- Tags: first-pair, litvoice, translation, rosetta, editorial-instrumentation

\[\](https://firstpair.org/books/rosetta/)

First Pair has published Rosetta: Comparative Voice Engineering as Editorial Method, our scholarly account of a research program that began with three voices for *Lighthouse Republics*, became the rights-aware LitVoice toolkit, and then became a Russian translation-review system for *The Invented Enemy*.

The article explains the origin of the South civic-documentary voice, the quieter East and sparer West experiments, the limits of author-name shorthand, and the scoring rails that let several complete editions remain readable at once. It also describes the Russian F/R/C Rosetta: Fable, Codex Review, and Codex aligned across 2,633 chunks, with separate measures for surface obstacles, textual divergence, corpus distance, equality, and human judgment.

Rosetta has now grown from a comparative reader into a complete editorial instrument. The current process preserves a simple rule: generated editions are evidence; editorial choices are a separate, source-bound record; accepted prose returns to the canonical manuscript only through a deliberate revision and rebuild.

## The current editing grammar

Every editable comparison exposes decisions at four scales without changing any displayed edition:

1. **Overall** chooses one pane for the complete aligned fragment.
2. **Sentence & words** reveals sentence-level preferences and a **Words** control for each nonempty unit.
3. **Whole sentence** selects the unit; word checkboxes select a contiguous phrase; **Only** selects one word directly.
4. The selected wording can be marked **Good**, **Bad**, or **Edit**.

Phrase records are independent. A sentence may contain several saved phrases, and a new phrase may overlap, nest inside, or cross an earlier one without erasing it. Clicking a saved edited highlight reopens its variants beneath the action row. **Save & replace** updates the preferred wording in place; **Save next** retains the earlier proposal and adds another. Stacked checkboxes select one preferred variant while preserving the alternatives as editorial evidence.

Headings participate in the same system. A **Heading** badge distinguishes a structural title from repeated navigation context, so a heading such as **Слово с двумя назначениями** can be selected, judged, and revised directly.

There is no generic “preserve previous choices” operation in the ordinary workflow. Reader pages are reproducible; decisions live in a separate three-file record: a machine-readable ledger, its integrity sidecar, and a readable digest. The record is bound to the project, exact source content, alignment contract, panes, sentences, tokens, and target spans. A mismatched or partially updated record fails closed instead of being guessed into a new edition.

## Reader and Review are different kinds of work

**Reader** keeps the alternatives in context. Its pinned navigation connects Previous, Up, Back, Top, Contents, Next, and the peer **Review** view. Desktop layouts preserve simultaneous comparison: the Russian vault gives long sentences full measure in vertically stacked F/R/C panes, while the Lighthouse vault keeps its configured edition columns and divergence rails. Mobile editions use horizontally snapping panes, direct edition tabs, and a visible next-pane edge. Preview vaults use the same control grammar on a bounded, inspectable corpus before it is applied to a complete editorial vault.

**Review** gathers only pages with decisions and turns those decisions into the appropriate next-stage artifact. The two projects intentionally do not pretend to have the same reconstruction semantics.

For *The Invented Enemy*, Review starts from Codex Review (R). An editor can leave most of R untouched, prefer another whole fragment or sentence where needed, and cherry-pick exact phrases from F, R, or C. **Bad** remains feedback and never enters the copy. A **Good** selection or the preferred **Edit** variant enters the candidate only after **Use in copy** maps it to an exact, non-overlapping destination in the currently effective base. Review then produces an inspectable R-first new-copy preview and can export Markdown plus JSON provenance. It never silently modifies F, R, or C.

For *Lighthouse Republics*, the aligned objects are conceptual lineages rather than a lossless manuscript stream. The same Overall, sentence, word, phrase, judgment, and variant controls therefore feed a deterministic next-edition instruction digest. Review links every instruction back to its fragment and pane, but does not fabricate a manuscript by concatenating lineages whose source order may split, fuse, repeat, or move.

In both profiles, the durable workflow is:

1. compare in Reader;
2. record decisions at the smallest useful scale;
3. inspect every touched page in Review;
4. export a candidate copy or instruction digest;
5. apply accepted changes to the canonical source;
6. rebuild the book, scores, EPUBs, and vaults under a new source identity.

That final boundary is central. Rosetta can make an editorial decision precise, portable, and reviewable. It does not make the decision disappear into automation.

## Read the research

The full article includes the *Lighthouse Republics* voice lineage, the creation of LitVoice, the Rosetta alignment and scoring model, Kindle-scale page renders, desktop and mobile vault views, the Russian corpus and N-gram work, claim verification, and the F/R/C translation case study.

Read *Rosetta: Comparative Voice Engineering as Editorial Method* on First Pair.

---

published with [omnighost](https://github.com/firstpair/omnighost?ref=firstpair.press) · SHA-256 e6fc844c373cad021316b4b07e7bb6ed26ab22849705b4133ae8b0257203f0da