Ask a language model to translate a novel chapter by chapter and the first chapters usually read well. The trouble shows up around chapter ten: a character called “the Keeper” becomes “the Guardian”, a nickname is translated once and kept in the original the next time, a captain loses her rank. Each passage, taken alone, is a reasonable translation. The book, taken whole, is not.
The cause is simple. The model only sees what it is given: the passage in front of it and, at best, a few lines of context. It does not remember what it decided three hundred pages earlier. Consistency does not come from a better model; it comes from giving every passage the decisions already made, and from checking afterwards that they were followed.
Here is the method we use in Libris Translate, and that you can apply with any tool.
Read the whole book before translating any of it
A name’s right translation often depends on what comes later: a title that turns out to be a family name, two nicknames that belong to the same person, a place that gives its name to an order. Translating the first chapter before reading the last means deciding blind.
Libris Translate therefore analyses every passage of a volume before translating the first one. The analysis produces a Book Bible, character sheets and glossary proposals, so the very first passage already receives the complete list of characters and terms. If passages are added to a volume later, their analysis is completed before translation resumes.
Turn each name into an explicit decision
For every character, place, institution or invented object, decide once: keep it, translate it, or adapt it. Then write that decision down where the translation will read it.
In Libris Translate, that place is the glossary. The analysis proposes terms; you correct the translation, then mark the term Accepted or Locked. “Add and lock” adds a term of your own. A locked term, for example “Keeper” → “gardien”, is sent with every passage where it appears.
Decide the names that carry the story first: those of the main characters, those that recur in every chapter, and those whose meaning matters (a “Tide Council” whose name the reader should understand).
Attach every alias to one person
A name is rarely alone. “Mara”, “Captain Mara” and “the captain” may be one person; “the old man” may be two. If the translation treats them as separate characters, it cannot keep their register, their gender or the way others address them.
The character sheets in Libris Translate hold a canonical name, aliases, pronouns, relations, speech style and translation notes. In the Characters tab, “Confirm same person” merges two identities the analysis suspects are one, and the other names stay as aliases. A similar name does not prove a shared identity: compare both sheets before merging.
Once a sheet is right, “Save and validate” it. A validated sheet stops learning: later analyses no longer add relations, aliases or a description to it.
Check the result, not just the instructions
Telling a model to use a term does not guarantee that it will. After translation, every locked term that appears in the source should appear in the target.
Libris Translate checks exactly that. When a passage’s source contains a locked term and its translation does not contain the locked rendering, the passage gets a finding “Locked translation missing”. It is an error. It becomes a warning only when the source may not be naming the term at all: a single Chinese or Japanese character that also sits inside other words, or a name made of letters alone that the source only uses at the start of a sentence, where it could be an ordinary word. The check reads the text itself: a passage cannot pass it because the model claimed it followed the glossary.
Change your mind without starting over
You will change a decision halfway through, and it is better to do so than to keep a bad name for three hundred pages.
When you save a term, Libris Translate marks the passages concerned for reevaluation, counts where the old word still appears in the passages already translated, and offers to replace it there without calling a model. In a series, a term accepted for the series applies to the following volumes unless a volume explicitly keeps its own translation.
What this method does not do
It does not make a translation literary. The glossary and the checks keep names stable; the style, the rhythm and the voice still need a reader. A consistent translation is the precondition for a good one, not its proof.
To go further: the guides Glossaries, Characters and identities and Analysis and translation describe each screen mentioned here.