I got you.
I guess there are two approaches. One is assuming the response for each block of those 4 TR's is linear, and use the behavior-weighted inputs for the block design. The advantage is that you have more degrees of freedom, but with the assumption of linearity in each block.
The other approach is to treat each TR in a block as a separate stimulus, and in the end you make different trend tests: linear, quadratic, or something else? In this case no assumption is made, but you would sacrifice with the loss of triple degrees of freedom.
Maybe try both?
Gang