Processing model

The pipeline separates translation generation from revision and assessment. This page describes each stage’s responsibilities and inputs; the document guide covers the operational controls.

Unit of work

A segment is the unit of execution, review and recovery. Segmentation allows part of a document to be processed independently, intermediate outputs to be retained, and a stage to be repeated without recalculating the entire text.

Each model request is assembled from the configuration and available data. The model does not automatically have access to the project’s history: context, terminology, examples and references must be included in the request.

Stage responsibilities

Stage	Main inputs	Expected output
Translation	Current source, document context, instructions and glossary	Initial translation
Refine	Source, context and previous draft	Complete revised translation
Format	Translated text and formatting instructions	Structural and Markdown syntax repairs
Judge	Source, translation and assessment criteria	Assessment and structured findings
Coherence	Translations and neighbouring translated segments	Findings about consistency across segments

Format uses a separate prompt. It does not receive the translation persona, glossary or source context. Its instructions limit it to formatting repairs, but the output still requires review.

In DeepL Hybrid mode, the initial stage uses the DeepL API and its language, formality and glossary parameters. Any subsequent stages use independently configured LLM providers.

Configuration and reproducibility

Separate stages make it possible to compare outputs and identify the step that introduced an error. They do not make generation deterministic: repeating a request may produce a different response, even at a low temperature. Valid JSON provides a parseable structure, not a guarantee of sound judgement.

To assess a change, keep the text and evaluation criteria fixed and change one parameter at a time. Use intermediate outputs and the operations console to compare the actual requests. The reviewer remains responsible for the final editorial decision.

Context and resources

The glossary specifies required terminology. Phrase memory supplies only the references selected for the current segment. Translation examples guide style across the pipeline.

See Context and prompt caching for request structure and prefix reuse, and Pipeline configuration for the available parameters.