I have tried to run the model and looked at the way the GRF are created, it looks ok.
In general GRF prediction only works is the motion is recorded realistically, this means that if the CoM of the model has unrealistic accelerations it will lead to muscle recruitment problems, simply because the model is about to tip over. These accelerations are necessarily visible but they may result in that model can not be balanced by the GRF predictions. You can look at the residual forces in the hip to get an idea it is also possible to increase the strength if the residual muscles which are working like a safety net in the model, they are weak so the model will try not to use them if possible, by increasing the strength of these (weakresiduals.any) it is likely to be able to handle more out of balance situations, but the problem is that with a higher strength the muscles will also kick in when not needed.
There are two paths of solutions:
Adjust the GRF settings so that the contact takes place in a longer time frame. This is done ForcePaltes.any file by adjusting the LimitDistHigh and friction coef. There is also a property named LimitVelHigh which can be adjusted. These adjustment should not be unrealistic… so there is limit to how much you can do here. From what i can see when the model fails the right foot has no contact, and it looks rather close to ground so it may help to increase the limitdisthigh
The other path to go down by is to apply your own force data, i can not say exactly how this is done without knowing more about the data, but in general what we do with the built in forceplate classes is to create a forceplate segment, drive it in space with no reactions, load it with the data availble, and finally create a reactionforce (no driver) between the plate and the limb in question.
Hope it makes sense and sorry for any typos.