There are always seeds; that how random number generator work. In the Geant4 toolkit, the same seeds are used at the start of the job, to allow for reproducibility and debugging. Users who run very large samples, with multiple jobs, will have their own method for setting unique, and different, seeds for each job to avoid generate the same events over and over again.
When you switch Geant4 versions, obviously there will be changes to some of the physic processes, which means that the detailed sequence of steps taken by a track will be different. That means that the sequence of random numbers generated over the course of the job will be different: the second event will start with a different random number in your G4 10.7 job than it did in your 10.6 job.
Thus, at the end of the day you’ll have numerically different results, but those results should be statistically equivalent. A very, very crude test of that is simply +/-sqrt(N). You wrote:
sqrt(1508) = 38.83. Both of your individual results are within +/-1 sigma of their average, so I don’t see an obvious problem.