I was really hoping for an retrospective about a multi-armed bandit learning algorithm that ran in production but never experimented, effectively yielding a "zero-armed bandit".