Partially observable models like POMDPs should know what the initial observation is when starting.
For state-based observations (O : S -> Z), this is straightforward because /observations/states/values.bin stores an observation for each state, including the initial state. These can also be stochastic.
But for action-state observations (O : A x S -> Z), this is less straightforward because UMB attaches them to branches s-a->s' (assuming that it will be consistent across different source state s for each action a and target state s'). When the initial state has no incoming transitions, there is no place to store an initial observation.
It may be sensible to consider this together with initial state distributions, which are also not yet supported.