One of the assumptions inherent in the RSS treatment of tolerances is that the quality level of all terms, input and output, is the same. If the part tolerances used as inputs to the RSS equation each represent 99.7% of the distribution of parts, then the Min and Max RSS resultant values also represent 99.7% of the predicted assembly-level variation.
However, I don't believe the TolAnalyst results are the true, traditional RSS results. Based on what I've observed the algorithm seems to be using a scaling factor and calculating what is sometimes referred to as a modified RSS, or MRSS. The factor can either be considered a factor of safety, like those used in stress calculations, or, if the inputs are assumed to be defined each with a Cp of 1.0, then the factor can be considered the Cp of the output. In other words, the calculated RSS limits would include a greater percentage of the assemblies then the input tolerances would include at a part level. Unfortunately despite my inquiries I've never been able to get a good description of the algorithm being used by TolAnalyst for the RSS results.
Thankyou very much Stephen