why does gnubg evaluate this position so wrongly

[Misja Alma]

I have the following position: 32UBAAfMzg2DAA
At the grandmaster level gnubg evaluates this as no double-beaver
(eq.: -0.355), while actually this is a borderline take/pass !
Now this is a crunch position, and I have been told that gnubg uses 3
separate neural nets, one of which is only for crunch positions. So I would
think gnubg would be good at this. Or do I have to do something special to
install the 3 neural nets?

My evaluation settings for cube decisions are grandmaster, lookahead 3
plies, no reduced evaluation, cubeful chequer evaluation, and 0.00 noise. I
am using gnubg build 08-11.

Ciao,
Misja