Tested on one of our ugc classification agents.

Tested on one of our ugc classification agents. 1st run showed the potential but also showed all the ways we were ambiguous in our prompt and just trusting llm decision making, after one iteration of the prompt the eval beat Luna by quite some margin