As a part of benchmarking, the model generated is run against multiple evaluation pipelines. The new has to outperform the older ones for it to be selected for further use.