There is quite a lot of theory behind the LOCAL algorithms (from Infrasoft International), anyway this is a first and quick approach to them.
The idea is to develop an equation for a certain constituent, based on the most adequate samples from a database (red file which has the math treatments already applied). Of course if there are not adequate samples we won´t get a prediction, in that case we can send that sample to the lab, and add the spectrum with its corresponding lab value to the database for future reprocessing.
During the calibration (on the fly) procedure, a certain number of factors, and a certain number of samples are chosen, always between the max. and min. that we have configured in the prediction model configuration.
There is not risk of overfitting, calibration is PLS.
I will include more post and videos about this interesting calibration method.