I think the best that can be hoped for at this state (both of machine learning and Go programming) is to identify a subset of atypical games for manual inspection. For example, games where a large group dies, or games lost against a weaker opponent.