Why an AI manga translator added a per-upload glossary to keep character names consistent

A developer building an AI manga translation tool found that users most often complained about character names and terms drifting across long volumes, such as a name changing spelling or an honorific being dropped. The problem arises because the model makes many small translation decisions across batched chunks, and Japanese romanization and inferred subjects add further variation. Earlier fixes like carrying forward summaries and enlarging chunks only partly helped, and post-processing find-and-replace was judged unsafe. The solution was a glossary attached to each upload that the translation step must respect, with CSV import and export so existing scanlation terminology sheets can be reused.