@alankilian
Any multi-axis measurement setup will have some form of ‘memory’ due to the loads last placed on it. When you have a repeatable load pattern (such as between x-axis points), you get fairly consistent results. When you have a different movement (move multiple axes back to the next line), you get different results.
The degree of measurement variability you get will depend on specific contributors of both the motion and measurement systems. Backlash can show as step variation, inertial loads that strain a force sensor can also show up. Creep, temperature drift, etc...
Without solving each of these issues independently, it’s easiest to randomize the motion (turning repeatable movement errors into nearly random errors) and perform more measurements. Might not work as well on the edges though due to less random motion on one side.
@dc42
IMO that’s the biggest flaw in the auto calibration at the moment. There are multiple ‘solutions’ from a calibration that tend to vary between calibrations. While each will produce decent prints, it should be obvious that things like endstop locations aren’t actually changing between runs.
Again, random motion and statistical modeling would likely produce the ‘real’ calibration. Not that it matters too much as the ‘fake’ one-time measurements are good enough to print.