It would depend on the level of complexity being discussed. This could be as simple as a grinder adjusting constantly to as complicated as a robot fed completely automated cell with 100% inline gaging.
If your complaint is 'the gaging system could be inadequate' a non compensating setup wouldn't work if the gaging was inadequate as well.
That's not true, because by constantly adjusting, you use less of the tolerance, and have a reduce PV causing your ndc to drop. Using a gage to make a constant adjustment requires more resolution that one that uses an acceptable amount of tolerance to adjust the process.
Even auto-compensating a grinder can be more complicated if one has tight tolerances and you are trying to meet tighter tolerances by an forcing your distribution to be "normal" and meet Cpk. A lot of unnecessary dabbling.
The process isn't a black box, you know the data you're feeding it and how it reacts to those changes.
If you don't know
how it reacts to the changes, it is black box. You might accept the black box output as acceptable, but that does not make it any less black box.
You have absolutely no idea what its logic is.
Is '(dimension) is now 0.0005mm above nominal, adjust associated tool 0.0005mm' a black box nobody can understand? You could even log these changes if somebody wanted to go back.
Even that logic is not a correct adjustment. You would adjust to 0.00025 below. Besides, do you adjust it based on its largest diameter, average diameter or minimum diameter?
Or do you throw all caution to the wind and ignore roundness altogether? That would be pretty darn sloppy for precision machining.
Logging the adjustment is part of my point stated above...although you would actually need to know what to do with the logged data to extract back out of it its value.
In essence, it sounds like to me as if you are supporting forcing an expensive, automated system to go back to the garbage statistics of X-bar/R and Cpk that have been shown to be incorrect and were corrected by the X hi/lo-R charting methodology and the correct capability calculation for the continuous uniform distribution. Doesn't seem to be a good idea to me in the long run. Hard to see the justification.