Most machine translation optimizes for meaning and drops everything else: tone, register, irony, the cultural weight a phrase carries. Artinya is an attempt to keep those, using a three-stage pipeline:
- Describe. Analyze the source text for style, tone, intent, cultural meaning, and symbolism.
- Translate. Translate with that description as context, so the model is working from an explicit account of what it needs to preserve.
- Evaluate. Score the output on accuracy, clarity, and preservation of style, and re-translate when it falls short.
The honest part
The README keeps a running list of things I am genuinely unsure about, and I think that list is the most valuable thing in the repository:
- How good is the translation, really?
- How expensive is it: how many API calls, how often does it re-translate?
- How often do the models fail to produce valid structured output?
- Does the whole pipeline actually beat a plain translation call?
That last one is the real experiment. A three-stage pipeline is more elaborate than a single prompt, and more elaborate is not the same as better. The ablation is the point of the project, not an afterthought to it, and I would rather publish the question than pretend I had already answered it.