The one random stream an episode draws from.

Monty threads a single mutable numpy.random.RandomState through every module (ctx.rng), so what a module draws depends on how many draws the modules before it made. That object is shared and mutable, which is why it is a process here: one per model, holding a Monty.Random.Numpy state, reseeded when the episode's seed changes.

Reached through Monty.Context.rng, a module draws in the order it is stepped. That order is Monty's — sensor modules, then learning modules' goals, then the motor system — and it is why the coordinator steps its sensor modules one at a time: a barrier would decide which module drew first by which call arrived first, and the numbers would follow the scheduler.

A draw is a call, so this is the one place in a step where a module is not a pure function of its inputs; what it stays is a function of the episode, and two runs of the same episode draw the same numbers in the same order.

Every call here waits as long as it takes. A step is computation-bound and the rest of the runtime already says so (:infinity on every server call); a draw was the one thing left on the default five seconds, so a step that queued behind a long one — or a suite running sixteen cases at once behind one stream — died of a timeout rather than finishing. A draw is not a health check: if the stream is slow, the step is slow.