Each sentence of a source book is translated into a target language by a pool of models — several general-purpose LLMs run at different temperatures, plus a dedicated machine-translation model. Each translation is then back-translated to English and compared to the original using cosine similarity on multilingual embeddings. A separate LLM, chosen only to judge and never to translate, blindly rates each candidate on grammar, naturalness, and fidelity. Combining both scores selects a winner for this first phase.

That winner then becomes the anchor for the next phase: self-refinement. The pipeline re-translates the current best translation, taking it as a hint and mutating it — an anchored mutation that keeps the sentence grammatical while searching for something better. This can repeat over several phases, each one anchored to the best result so far. The final winner is chosen across all phases played, so the finished document is a hybrid not of models but of phases: the best translation for each sentence, whichever model or phase produced it.