Dear Hans,
If you refer to my particular case, the negative weight was explicitly set.
The thing that made it tricky to find was that the error occurred during
the simulation and not when the model was created (which is the part
where the problematic synaptic connections were established).
Best
Angeliki