Not sure if this is a bug or this just comes with the territory (i.e. numerics). When I solve this nasty MINLP with Baron using an allowed gap optcr=0 I get:
Solution = 0.922492168619502 found at node 6
This is actually very good performance. However there is better solution (we are maximizing). If I plug this solution into the model I see:
Solution = 0.954564813873509 best solution found during preprocessing
This is somewhat better (about 3% improvement). This looks like a little bit more than expected.
The model is related to some reliability problem: