Well I think I am seeing it in a test I am running (I am using very wide ranges).I am not worried by this case. I never saw it happen in practice, and, if it does happen, it is not very difficult to figure it out by looking at the data.
I suspect that in my test the evaluation function has two local maxima and its seems the quadratic logistic approximation by CLOP has a minimum between the local maxima (so the Hessian is not positive definite). It is a bit like in the above picture by Kai Laskos.
The average of the samples falls between the local maxima. I suspect close to
the local minimum of the quadratic logistic approximation.
I am currently at 150000 games. I'll let it run till 300000 or so to see if the
the Hessian becomes positive definite (in that case my assessment of the situation would
be incorrect).