A board heats up, a sensor samples it every 100 ms, the last four samples are averaged, and an LED turns on above 80 °C. Each tab runs it under a different model of computation.
The physics: the board's temperature exists at every instant. A computer can only approximate it, by stepping an ODE solver forward. Change the step size and watch the approximation drift from the true curve.
Every message is an event with a timestamp, held in one global queue. The director always processes the earliest event, and model time jumps from one event to the next: nothing in between is simulated.
Sorted by timestamp, then microstep (μ), then depth: upstream actors first. All events here have μ = 0; microsteps only grow when an actor reschedules itself at the same instant.
There is no clock. An actor may fire whenever enough tokens wait on its inputs. You are the scheduler: click any green actor, in any order you like.
Averager consumes 4 tokens per firing and produces 1, so a valid schedule fires the sensor 4 times for every averager firing. Reads are destructive: consumed tokens disappear. Unlike the sliding window in the DE tab, this averager takes non-overlapping blocks of 4: a multirate actor, which is where SDF rates come from.