How are you supposed to use MaskSplitterNetwork?

[profPlum]

I've tried using mask splitter network like so:

# This is the secret sauce to make Actor & Value Network accept dict inputs 
def get_dict_combiner():
    combiner_seq = tf_agents.networks.Sequential([tf_agents.networks.NestFlatten(), tf.keras.layers.Concatenate(axis=-1)])
    combiner = tf.keras.layers.Lambda(lambda x: combiner_seq(x)[0])
    return combiner
    
actor_net = actor_distribution_network.ActorDistributionNetwork(input_spec, train_tf_env.action_spec(), preprocessing_combiner=get_dict_combiner(), fc_layer_params=(256,128,64))
value_net = value_network.ValueNetwork(input_spec, preprocessing_combiner=get_dict_combiner(), fc_layer_params=(256,128,64))

def splitter_fn(obs_dict): # NOTE: don't modify obs_dict it will cause bugs!
    return {k:v for k,v in obs_dict.items() if k!='possible_moves_mask'}, obs_dict['possible_moves_mask']
actor_net=MaskSplitterNetwork(splitter_fn, actor_net, passthrough_mask=True)
value_net=MaskSplitterNetwork(splitter_fn, value_net, passthrough_mask=False)

Yet it appears that sometimes the actor is still guessing actions which are disallowed by the 'possible_moves_mask'. How is this possible? Do I need to explicitly use the mask to prevent guessing of invalid moves somewhere? I was under the impression this happened automatically right?