RuneTranslate

Neither exactly. Enforcement doesn't go through the prompt at all: glossary terms get swapped for placeholders before the line reaches the model and restored afterward, so the model can't drift on them.

For new terms, there's an auto-detect pass that scans the whole project before translating, picks up recurring katakana names, speaker names and repeated kanji compounds, and suggests them for the glossary with a romaji guess you can edit. So the idea is to catch them before the model ever gets a chance to invent three versions. What it doesn't do yet is audit the output afterward for terms that came out inconsistently. That's a good call though, I'll probably add it.