Since its origin, the navigation services provided by the GNSS system have been applied to a variety of field, some of which are beyond the original scope of the infrastructure. LEO satellites autonomous orbit determination by means of on board GNSS receiver is one of the cases. The achievable position accuracy, in the order of 10 m 3d rms, satisfies most of the mission requirements and can be implemented with a simple navigation algorithm based on loose constraining of observations. For those missions that require better real time accuracies, sub meter positioning is achievable, by means of reduced dynamic filtering, which provides strong constraining of observations thanks to a detailed orbit force model. The algorithm general architecture is well covered in literature, although its real world implementation is still limited to few cases, partially because few are the missions that require such accuracy in real time, and partially because of its implementation complexity. This paper addresses one aspect of this complexity, which is related to filter tuning, or in other words to the effective definition of a statistical description of observations and process errors. Being the filter tuning a process of trial and error in most cases, and a delicate implementation step at the same time which may lead to instability and divergence, this paper proposes a procedure to achieve stable sub meter positioning from the reduced dynamic algorithm implementation.