A little real world experiment.
I set out nails at ±300' intervals along a half mile of road and levelled through them, up and back, using a Leica DNA10 digital level. Loop closure was flat - zero.
Tied all the nails using a Leica GS18 base/rover pair. Employed the tilt sensing feature on the rover, so none of the shots had a plumb pole. GPS base was between a half mile and several hundred feet of the tied points. Base location had good, but not great, sky. I held 2 of the levelled points for elevation.
The road where the nails are is mostly tree lined and some of the locations were very poor for RTK. None of them are really what you might call good. Each RTK result is an average of at least two 3 second occupations. Weighted by the dc for coordinate quality.
So you see that the elevations are fairly tight where the conditions approach good, and fall off where conditions are really bad.