[Discuss-gnuradio] QPSK demodulation: Constellation Soft Decoder question

[jmfriedt]

I am investigating Meteor M2 decoding. So far I had tackled BPSK
decoding (RDS, GPS) and it is the first time I analyze QPSK signals.
Based on github.com/otti-soft/meteor-m2-lrpt.git, decoding works
fantastically well, but I don't understand why. Costas and MM Clock
recovery are understood, but it is the first time I meet the
Constellation Soft Decoder. I am unable to find any information on its
output and operating principle, but most worrying to me is that I
believed that the real output would exhibit four levels following the
decision of which bit pair was the most probable, but when displaying
on a scope the output of the Constellation Soft Decoder I only see two
well defined levels.
Can anyone bring some insight on this processing block, which would
help me understand the structure of the resulting file and hence start
tackling the digital decoding part of the protocol. 
The flowgraph I use in this investigation is visible at
http://jmfriedt.sequanux.org/airspy_m2_lrpt_rxtest.png with an input
file containing complex I/Q samples collected during a Meteor M2 pass.

Thanks, JM