CatScribe Docs

#AI Translation Consistency

Consistency is a systems problem, not a single-model setting. No model — neural MT or LLM — natively remembers the decision it made 200 pages ago. Every long translation project therefore needs external machinery that records decisions and enforces them on every subsequent segment. This guide explains that machinery and how CatScribe implements it.

#Why Models Drift

Translation systems process text in windows: a chunk, a context window, a batch. Anything outside the window does not exist for the model. Consequences:

  • The same term is translated plausibly-but-differently each time it appears, because each occurrence is a fresh decision.
  • Synonym variety — a virtue in monolingual writing — becomes a defect: "the Order", "the Brotherhood", and "the Society" for one organization.
  • Register drifts as surrounding text changes: formal address in one chapter, informal in the next.

Drift is not a model bug to wait out. It is structural, and the fix is architectural: a decision store (glossary), enforcement at translation time, and verification afterward.

#The Glossary As A Contract

Treat the glossary not as a dictionary but as a contract: this source term is always rendered as this target term. Rules for a glossary that actually enforces:

  • One canonical target per term. If two renderings are genuinely acceptable, pick one anyway.
  • Cover the reader-visible surface: names, places, organizations, titles, honorifics, invented vocabulary, recurring slogans.
  • Record inflection information where the target language needs it — CatScribe entries carry optional gender, number, and variant fields, plus an "Allow inflection variants" flag, so matching survives grammatical inflection.
  • Update it the moment a decision is made during review, not at the end of the pass. A glossary updated late enforces the wrong half of the project.

#Translation Memory: Consistency Beyond Terms

Terms are the enforceable part of consistency; phrasing is the rest. A translation memory stores whole approved segments so that recurring or similar sentences are rendered the way they were rendered before.

CatScribe's memory has a deliberate property worth copying in any workflow: only human-approved segments enter the memory. Machine output never trains the system on its own guesses — approval in the CAT Editor is the gate. Stored segments are then supplied as consistency context to AI refinement stages in later work. The practical rule: approve as you review, because every approval compounds.

#Multi-Pass Refinement

Trying to fix accuracy, terminology, and style simultaneously produces churn: a style edit reintroduces a terminology error, whose fix breaks the style. Sequence the passes instead:

  1. Pass 1 — baseline: complete machine draft; fix failures and gross meaning errors.
  2. Pass 2 — terminology: enforce the glossary everywhere; bulk-fix with find-and-replace; update the glossary with every new decision.
  3. Pass 3 — style: harmonize voice and fluency, using AI refinement where useful, inside the now-stable terminology.

Each pass is a quality gate: do not start style work until terminology stops changing.

#Measuring Consistency

What you cannot measure, you cannot finish. Useful signals CatScribe exposes:

  • Glossary success rate per chunk (matched/total terms) in QA flags and the score tooltip.
  • Chunk quality scores ("Score = (2×Quality + Glossary) / 3") — a low-score queue is a prioritized review list.
  • Review progress (approved/total) — approval percentage is the honest definition of "done".

#Why This Works

Layered enforcement converts consistency from a per-sentence hope into a system property: decisions are recorded once (glossary, approved segments), applied automatically everywhere (shielding, prompts, memory), and audited mechanically (verification, flags, scores). The human's job shrinks to making each decision once — which is the only part that ever needed a human.