How do you adjudicate the request to build a new attrib...

Building a new judge from scratch duplicates the complex work of training a model to discriminate between silence and contradiction. Instead, reuse the existing fact-checking judge by mapping the labels to the new vocabulary. However, you must address the difference in evidence sources: fact-checking typically uses a curated, static corpus, whereas RAG uses dynamic retrieval. This shift in class base rates necessitates separate calibration for the shared judge to ensure performance remains stable.