Every variant sees the same arrivals and the same drivers for a given seed (common random numbers), so differences come from the warning strategy only.
Driver behaviour parameters are stated assumptions (see docs/ASSUMPTIONS.md) and are varied below.
Results are relative comparisons under those assumptions — not a prediction of real crash reductions, which needs a field pilot.
If QueueGuard only worked under optimistic assumptions it would not be worth piloting. Each chart varies one assumption and keeps the rest at the defaults.
Chance a driver notices a warning
Distance at which unwarned drivers recognise stopped traffic
Peak demand (longer queues)
Warning placed too late (real tail within 150 m of the warning point 90 s later)