Nice read. How did you manage the increased compute cost of moving from sampled to full scale runs? I imagine the parallel runs could be potentially prohibitive if the full scale pipeline requires a lot of compute