Gravar-mail: Stable Learning in Stochastic Network States