I think Yang Chen has experimented with the idea of tampering random number sequences generated by Csmith. What he found, IIRC, is that Csmith makes hundreds of random choices (each of them determined by a random number within a range, or a flip of true/false), if we change just one random number in the sequence, all the subsequent random numbers are affected, and the generated program looks vastly different from the original one.
I think the alteration of the random choices has to be done after, not during, the program is generated.
-Xuejun