Radiance uses “stochastic” methods with inherent randomness that causes variations between simulations. This is a feature, not a bug! Without these methods, simulations would take far too long to be of any use.
These variations can be particularly apparent for metrics that bin data (e.g. the 50% of daylight hours above 300 lux for sDA calculations). Values very near cutoffs can fluctuate between bins between simulations - e.g. 49.6% of hours vs 50.2% of hours for a point in an sDA score.