Let me rephrase what BevD said because it could be interpreted multiple ways.
There are three possible scenarios for gage application(s):
- Part inspection only
- Statistical uses (i.e., Statistical Process Control (SPC), capability studies, DOEs, etc.) only
- Both 1 and 2
For scenario 1, part selection is discretionary because the %Tol metric does not use the study variation as part of the calculation. If linearity studies have or will been done, it truly does not matter how the parts are selected. If linearity studies have/will not be done, it is a good practice to spread the samples out over the full range of tolerance and even beyond.
For scenario 2 (and 3) part selection is absolutely critical if %SV (Study Variation) or ndc (number of discrete categories) will be used. The parts absolutely MUST reflect actual process variation. If the parts are deliberately selected to be a larger spread than the actual process, study variation will be inflated and the %SV or ndc will be very misleading. Doing such with full knowledge of the implications is unethical. A better approach is to use the results of a capability study to calculate the %PV (Process Variation) metric. %PV is more accurate than %SV.