Just a thought: could we unleash a deep learning agent into an environment that codifies all the rules of a given law/policy system (e.g. tax), and with an appropriate reward system, expose possible loopholes in the system more quickly?