At Applied Compute most of our RL and eval pipelines grade with an LLM judge. These produce one result per rubric criterion, typically a paragraph of text accompanying a score. For some complicated tasks, grading can be a slow, pricey part of the loop.

We swapped a grader on a