Fixing Broken Markdown in AI Translation

This post came from a practical failure, not an abstract design exercise.

Community feedback around Japanese translations showed me that many AI translation problems were not only semantic. A translation could sound reasonable and still leave the document broken. Anchors, code blocks, lists, links, frontmatter, and formatting structure could drift just enough to make the result hard to review or unsafe to merge.

That changed how I thought about Co-op Translator. The output was not just text. It was a software artifact that needed structure, state, validation, and recovery paths.

The article explains that shift: a translation pipeline has to protect Markdown as Markdown, not only produce fluent sentences. If maintainers cannot review the diff with confidence, the generated content has not really succeeded.

The lesson is that generated documentation earns trust through the system around it. The model matters, but the pipeline decides whether maintainers can actually use the result.

Read the original article:

Fixing Broken Markdown in AI Translation: Hardening a Production Pipeline