The best the games can do is create a pseudo-random sequence, which will be exactly the same on every power up since the exact same code path is taken. Input timing can be used to seed the PRNG, but if no buttons are pressed, every run will be the same.
It should be easy enough to debug the code in MESS and see how the PRNG is seeded, and that may help figure out the difference between the two.
I hacked the Robotron ROMs to always read the player 2 start and move up inputs as set, then turned on both of my games. Since there was no user input to give timing differences to the PRNG, the 2 games played identically. But the same ROM in MAME plays differently since there isn't perfect low-level emulation of the entire Williams system.
In this case, I bet they are using an uninitialized register or RAM location to seed the PRNG, and it turns out that it's always powering up the same, but MESS is initializing it to a different value.
