In this MR:
RANMAR inside tauola for each and every event is initialized with the same seeds at https://gitlab.cern.ch/atlas/athena/-/blob/master/Generators/Tauolapp_i/src/TauolaPP.cxx#L160 Which means that taus in events with the same total numbers of taus per event will be decayed in the same way. In their CMS, of course. To fix this behaviour the generator should be initialized once. Note: I don't know how many years it is there.
According to the docs in taoula the random seed should be 0<=IJKLIN<=900 OOO OOO
C ENTRY RMARIN(IJKLIN, NTOTIN,NTOT2N) C Initializing routine for RANMAR, may be called before C generating pseudorandom numbers with RANMAR. The input C values should be in the ranges: 0<=IJKLIN<=900 OOO OOO C 0<=NTOTIN<=999 999 999 C 0<=NTOT2N<<999 999 999! C To get the standard values in Marsaglia-s paper, IJKLIN=54217137 C NTOTIN,NTOT2N=0