This process seems back-to-front. Surely the best approach would be to AI review the papers immediately using multiple LLMs giving the paper ratings out of 10 in multiple categories. Then the "top" papers get to move on to the next stage, ie. human review