So if they are not sealing, what is causing the change in observed contact levels?

You can not simply model observed 'symptoms' and hope to get out of it a predictive model, you must have some idea of what is physically going on.

Were the logs taken after some production, so could the observed water levels be from a high perm stringer sucking water from the flanks?
Are faults blocks actually sealing?
Capillary pressure changes with perm/facies change?

Come back with your idea of what is actually going on, and I'll help suggest an appropriate way to initialize a model to honour it..... otherwise I'm sorry but can be of no further assistance