As discussed in other papers in this volume, the IAU XXIII General Assembly adopted a new fundamental celestial reference frame: the International Celestial Reference Frame (ICRF) based on VLBI observations of extragalactic radio sources (Ma et al., 1997). It is approximately 300 times more accurate than its predecessor, the FK5. At present, no other technique has produced a more accurate celestial frame than VLBI, Since no other astrometric technique provides an external standard of accuracy, the VLBI claim of a great leap forward in accuracy must be verified by internal consistency tests. This paper addresses one aspect of internal consistency: the ability of independent VLBI software packages to reproduce a celestial frame without significant loss of accuracy. This is no small task since the software packages are large - involving on the order of 100 000 lines of code. What does VLBI software do? Aside from routines designed to collect the data and extract raw observables which will not be considered here, its principal task is to model the differential group delay and phase delay rate of radio signals received at two widely separated antennas (Sovers, Fanselow & Jacobs, 1998). The software then refines this model via a least squares adjustment of relevant physical parameters which describe station locations, source positions, clock offsets, atmospheric refraction, tidal effects, etc. In the early 1990s, studies revealed that differences in software implementation and analyst’s choices of model options were one of the largest contributors to differences in independent calculations of VLBI celestial frames. These differences were of comparable size to the formal uncertainties of the celestial frame’s source positions.