Thank you very much ! :) The idea was indeed to not rely on LLM to watch LLMs, also making it pure stats allow ultra fast feedback loop. I tried a lots of different techniques, but well refined stats outperforms all neural network techniques