Castia vs. DeepL.

What DeepL actually does

By DeepL's own description of its technology, it translates using neural network architectures - the company describes combining a neural MT engine with a specialized large language model, evaluating whole sentences for context rather than translating word-by-word. That contextual, whole-sentence approach is a real strength over older phrase-based systems, and it's a large part of why DeepL is widely regarded as producing more natural sentence-level output than earlier machine translation.

What DeepL's own published description doesn't include is a step where the system generates multiple differently-angled drafts of a line and has a separate stage score them against a rubric before choosing one - it's built to produce one strong translation per request, which is the right shape for the documents and general text it's aimed at.

Why a lyric or a character's line needs more than one draft

A single, even very fluent, translation of a lyric still has to make a choice: keep the literal meaning, or bend it to preserve rhyme and rhythm. Peter Low's "Pentathlon Principle" treats that as an unavoidable tradeoff between five demands (singability, sense, naturalness, rhythm, rhyme) rather than something one correct answer resolves. Castia's Creative Adapter stage exists to actually make that tradeoff on purpose, five different ways, instead of settling on whichever balance one pass happened to land on.

And because a creative rewrite can drift from what a line actually means if nothing checks it, Castia's Judge stage scores every rewrite against a literal anchor translation before picking one - the same "have a model score candidate outputs against a rubric" approach NeurIPS 2023 research on LLM-as-a-judge evaluation found can track human preference well, while also documenting real biases (verbosity, position) that a judge design has to account for rather than assume away.

Side by side