Yes that’s exactly what I saw. Accuracy with bs=32 went from 97%->90% IIRC.
It was my intent to have the next iteration of calcs be part of the graph - I did it this way because I was trying to avoid having a continuously growing history in the graph. Which I think is working AFAICT.
But it’s certainly true that I don’t fully understand the details of all this, and I’m sure there’s things that need to be fixed to make this work with the “occasional skipping”.