The shape of a matatu network. Somebody drew those routes on a wall before anybody modelled them, and the graph has been running in production ever since. I want to know whether it is close to optimal or merely old. Those are different questions and only one of them has a clean answer.
Whether a model can think in Kiswahili without translating first — and what first would even mean there. Every benchmark I have seen measures the translation, not the thinking.
Queueing at a bank on the first of the month. The arrival process is not Poisson, everybody who has ever built the system knows it, and the model in the slide deck still says Poisson because that is the one with a closed form.
Why so much infrastructure work is deciding what to throw away. Deletion is the design. Everything else is accumulation with a schedule.