it's an attempt to vary the weightings of the information from the current versus previous step, to better adapt the Adams-Bashforth solver to work on non-uniform step sizes (exhibited especially by the "native" schedule). if the previous step was bigger (in log-sigma space) than the ...