Hi,
Thanks. Since observed is in [0,1], I tried adding the potentials
(P1-P2)*observed
or
(P1>P2)*observed
but I have the feeling this is completely wrong (Iām a newbie) 
Could you please point me to some paper about the likelihood-free inference approach you mentioned?
In the past I helped developing something called Probabilistic Constraint Programming, I wonder if it is related and how.
Thanks
Marco