Are we testing the model or the harness? If benchmarks show certain numbers it should perform as such without it