> Is there a reason why these long ISI's are, in fact, related optimally to
> power? The only thing I can think of is that sticking a couple of long ones
> in a run decreases expectancy effects and thus expectancy-related noise.
> Is RSFgen designed with such cognitive processes in mind?
No, you've overestimated the IQ of RSFgen. The occurrences of long ISI's are there simply because their presence makes the regressors less correlated with each other, thus leading to higher efficiency/power.
> If BOLD is roughly constant after 20 seconds post stimulus, can I just
> shorten the long ISI's to that amount?
An option could be added in RSFgen to set up an upper bound in ISI in the future. In reality RSFgen is more important in avoiding the worst designs than most efficient ones. So you may just check those designs in the top list and see if you can find one with all ISI's less than 20 sec.
Gang