Comment by kovezd

By the composition of evals. Plus secondary metrics like parameter size, and token cost.

Not perfect, but useful.