Its kind of elaborate but you could write an expression that grabs the mask vertexes of the mask that is creating the semi transparent layer and then automatically mask an adjustment layer (with a dummy mask already on board) using those coordinates. The blur would always be there on the adjustment layers and only shine through once a mask was positioned.
There are a few hoops to jump through to get access to the mask points though.
Better idea. Do you whole project without any blur. Then duplicate the comp and bring in that second copy to the first. Inside the second copy turn off all layers except the layers making the transparent layer.
Now back in your first comp use that second comp as the masking layer for a blurring adjustment layer.