Rerankers trained on peer review → paper pairs from the 44B corpus, and 44b Eval, the frozen 9-way hard-negative benchmark they are measured on.