The real world sanity checks for cheating make sense. But the nature of the regular tests can't be secret. Carmakers need to know the bar they are supposed to hit. Building something on the speculation that it might pass, maybe, is a whole different business model.
"Build a car that doesn't emit more than X of exhaust Y when a typical driver drives it a) ten kilometers in the city, b) a hundred kilometers on the highway."
That seems like a good enough requirement to me. Define "a typical driver" as "out of a hundred random test drivers, no more than 20 exceed the thresholds, no more than 5 exceed the thresholds by more than a factor two." That forces the car manufacturers to have a sufficient safety margin in their emissions.
I don't know how difficult the actual measurement is, but maybe you could pay a couple thousand people a reasonable amount of money to have some devices attached to their cars for a month or two and collect data. Or make the car makers pay for the procedure.
So this is going to happen once they've built the car + engine, and started selling it? Is 'car tester' going to become a side-hustle?
I agree that a degree of randomness is likely a good idea to avoid defeat devices, but one also has to consider that it could have two unintended consequences: 1) more expensive cars, due to more stringent QA procedures 2) relaxing of standards to ensure that companies can still practically make cars that conform to 'standards.'
As ever, it's important to consider that a layman's "seems reasonable to me" is another experts "that's not how things work."
Driving style has a profound effect on emissions. As does ambient temperature, uphill vs downhill, time at stoplight. Short trip vs longer (catalytic converters don't work till they are hot).
I'm sure they can find some way to address the cheating, but it's likely not the case that they are just missing some obvious simple solution.
Modern cars have enough sensors that they could store all necessary information about trips, have the service center read it out each time the car is serviced, anonymise, and send back to the manufacturer. That way the manufacturer would have accurate, real-world data about their 'average driver'. Could even make it country-specific.
These tests, though, are the ones you have to pass before you can sell that model of car, administered by government agencies. The agencies dictate the tests and measures.
Are you a software developer? Could you write the code for an application specified that loosely?
(At this point, every one of us is saying, yeah, I have. It just took a long time, a lot of change requests, and many round trips with a customer. Which is the point.)
Would it be feasible to have two tests? One public, and one where the details are hidden. The hidden test has a significantly lower pass bar, so that the only way you would fail it whilst passing the public test is if you are manipulating the outcome of the public test.
OK, but there are test details that are known to the manufacturers but seem irrelevant to the thing being tested. Like the bit about the angle of the steering wheel. Surely there are multiple ways one can test for the same standard?
The cheats are based around detecting when a car is strapped down to a dynamometer, indoors. That why the cheaters are looking at ambient temp, steering wheel angle, etc.
