[btwillard]

With our indexing approach, it only costs a dictionary lookup to get the next valid tokens during each sampling step, so very little latency.