It turns out that my simulations were running about 5x longer than necessary.
It wasn't wrong, but the system reached steady state after about 50k steps and I was running each one for 250k steps.
So instead of only doing one more batch I get to do more batches and hope this gives good data.