Take a random sample of machines and compare a number of pretend-votes to the outcome. But if cars change their behavior based on detecting a test environment,

Take a random sample of machines and compare a number of pretend-votes to the outcome. But if cars change their behavior based on detecting a test environment, how would you trust a voting machine to not do something similar? In the end a trusted minor miscount would be less bad than a count that is widely untrusted despite being correct: one would mean living through a "what if some people would have cast a different vote" alternative reality, the other would be a full blown legitimacy crisis, with unconstrained cheating in the next election as the least bad of all possible outcome. Just stick to paper ballots.