In the doc I believe it’s explained that the machine is just a wrapper for the model (a container with just the hyperparameters) and essentially the “fitresults” (whatever is learned during the fitting process that needs to be passed on to the predict or transform method).
So in this case if you want to recuperate the actual trained tree, please have a look at tree_machine.fitresult; IIRC the first element of the tuple is the tree that you want. This is also returned maybe more transparently by fitted_params(tree_machine) as a namedtuple.