Leah,
> Are the AFNI ACF blur estimates supposed to be using stats.s####_REML file as the final dataset?
Conceptually ACF/smoothness estimates should reflect the spatial relatedness in pure "noise" because you're supposed to control "false positives" assuming there is no signal in the data under the traditional statistical framework. So, it may make more sense to use the model residuals at the subject or population level to estimate the ACF parameters.
Gang