This one inspired me a bit:
So I changed the model to be built with logical blocks, but still have to use memory blocks.
With step size and solver set to auto, both versions (in one model, chose via manual variant sink) create different output. When the maximum step size is reduced, output becomes similar.