A rented humanoid does not arrive knowing your task, and the single biggest reason a trial disappoints is that somebody expected it to.
What arrives is the platform with the manufacturer capabilities working: it walks or drives, balances, maps a space, recognises and grips objects it has been shown, responds to voice and to a handheld remote, and can be teleoperated by a person wearing a headset. On top of that sits whatever we program for the one task you and we agreed before delivery. That task is the trial. It is deliberately one task, because a fortnight spent on one produces a number you can put in a business case, and a fortnight spent on five produces an anecdote.
Engineering time inside the rental is real but finite, and it is the part worth negotiating rather than the weekly rate. A two-day event needs almost none: the unit greets people and the stock behaviour is the product. A process trial needs several days of application work before any measurement starts, and if those days are not in the agreement the trial ends in the week it should have begun. Teaching is different again: the platform ships with a published SDK and ROS 2, and the point is that your students write the code, not us.
What a trial will not settle is reliability over months. Two weeks tells you whether a task is technically achievable, what the cycle time looks like and how the people around it react. Mean time between failures, wear on a gripper and the cost of a year of spares are not observable in that window, and anyone offering to prove them inside it is selling something.