Public control points constitute an integral part of the Japan's geodetic reference frame. Static GNSS surveys, directly referenced to multiple nearby stations within the nationwide network of continuously operating reference stations, have become the mainstream for second- and third-class control surveys. Coordinates are determined by three-dimensional network analysis of GNSS baseline solutions, which are eventually weighted independently for the north-south, east-west and up-down components by the inverse of their respective constant variances. Because the baseline distances of these baseline solutions typically vary from a few hundred meters to tens of kilometers, the distance dependence of the baseline solution uncertainty must be clearly defined. The quality of GNSS baseline solutions obtained in accordance with the General Standard of Operation Specifications for Public Survey (GSOS) has been systematically evaluated over a distance range from a few hundred meters to about 40 km from actual verification data of GNSS instruments used in public surveys. The root-mean-square errors (RMSEs) in north-south, east-west and up-down components relative to the reference values at GNSS instrumental verification sites clearly increase with baseline distance. The RMSEs in north-south and east-west components show statistically insignificant correlation, and exhibit a rather isotropic distribution in the horizontal plane. The RMSEs in each component are properly modeled by simple linear regressions on the baseline length and the same model can be applied to the north-south and east-west components. Furthermore, the square of the model value at that baseline length is found to give the variance of the baseline solution for each component. Because the inverse of such model variances plays an essential role as standard weights in network adjustment of baseline solutions, we compare the baseline distance-dependent model with the GSOS-prescribed constant variance model and variance-covariance matrix of the baseline solution when applied to the adjustment of experimental GNSS observations relevant to second- and third-class control surveys. Results show that the distance-dependent model and matrix perform equally well, and the relative height results between the new control points are consistent with total station measurements. The constant model, however, shows significant deviations from these results, which is a critical issue in geodetic discussions.