Why my thesis evaluation pipeline scored every run with both a deterministic benchmark and an LLM judge. The disagreements between them turned out to be the useful part.
MTEB scores won't help you here. How I choose embeddings when the data has structure.