Hi,
It seems to me that in order to justify the use of proportional weighting you have to determine whether you expect that incorrect responses activate the same brain regions as correct responses but in a proportional manner (more or less). If you expect different brain areas to be activated by correct and incorrect responses then you probably want to use 6 separate regressors in 3dDeconvolve in order to examine differences between correct and incorrect trial types across conditions.
Assuming proportionality presumes that a particular brain area is expected to be active in both conditions in some sort of ratio, which is fine if an area is not inactive in one of the conditions (i.e., zero activation). The fact that correct and incorrect trials are unequal in quantity means you will get a noisier fit for those conditions, but you can still legitimately compare them directly.
-jim