I wrote a custom calibration procedure (in JavaScript) and could not get above 2 stars - for all those who have a similar problem:
i solved that issue, by
giving the user time to rest his eyes for ~500ms before starting to collect data (sending 'pointstart') in every point. and not - like before - jump from one point to another - that gave me
huge errors - like 15 degrees and such. - this could be mentioned in the calibration api to help former developers, maybe.
now i easily get 5 stars with my own procedure
and implemented your star-ranking, too - thank you for pointing that out to me.