The hidden accuracy cost of batching LLM judge calls

I worked on a project that required tracking themes across weekly windows of incoming data. "Theme" here means a claim, not a tag or a label. Incoming text is de-noised and fed to an LLM to extract th