Gysbreght
I could not let his reply pass, because what he wrote in post #73 is absolute rubbish.
Just to keep this pot boiling, some comments to try and back the position as stated in post #73! I will assume that the main issue is whether an exact match of a model to the data is good or bad.
I can only speak from my experience and training in analysing physical data. All data is noisy and dealing with the errors is a big part of the data interpretation. There is always the risk of developing a complex physical model with lots of variables that the analyst is trying to get out of the data, e.g. in the case of a plasma, the temperature (sometimes several), density and chemical composition. If there are only a few data points, the model can be so complex that almost any data, including the noise, can be fitted by changing the variables to some set of values. If this happens, then the model is too complex and very little can be extracted from the result. So if any model can exactly match the data values as measured (assuming there is more than one data value), the analyst should be suspicious. One of the major reason for having a model is to be able to recognise the errors in the data and deal with them.
So the predictions of the model, when compared to the data, must still show a disparity/difference that is consistent with the known errors in the data.
In my experience it is usual to keep the comparison of the data and the model in the frame of the data where the errors are well understood. If others wish to do it another way, that’s fine, but the errors have to be transferred in some fashion because they have not gone away.