No need to decode if you can call API methods, receiving proper replies. You start learning (reverse engineering?) the subject's behavior based on responses to methods.

OK, I'm thinking as I type. The API must represent a functional system (i.e. function calls do no have side effects, produce the same output given the same environment) for RE to work. The first million iterations of such an API would not 'feel' functional because humans tend away from rationality (insert gender-prejudiced joke here.)

I suspect that once you get down into the tiniest bits of the human system (quantum behavior in brains, if indeed that's even necessary for consciousness), you can get much more functional results. Then, that gets us into eliminating chaos (replacing it with knowledge) in complex systems.

Consider: knowing how that precise collection of molecules (that tasty roasted meat, whose chicken was raised on marigold petals but once ate a bug or three) will affect the digestive, circulatory and nervous systems of a human; initially when the odor affects the receptors in the nose and ultimately after digestion has done its job and delivered all the tiny chicken bits to cells in the human body. And between those times, how anticipation from delayed gratification ("we'll wait for Grandma to arrive before we eat") affects the effects of the food on the human. And from here, chaos theory...

I pontificate muchly and pointlessly, but it was fun.