Presumably they could each create a graph for their own models and internal datasets, then use public benchmarks to calibrate between them.