Hardware level trust would be really cool. It seems theoretically possible because all you need is an accurate measurement of some physical property that is hard to forge when changing transistors orders etc. Practically though, anything like that would be most likely affected by manufacturing tolerances for the transistors, so you’d have to find something that allows a certain amount of error when individual transistors change, but will reveal unknown transistors and connections.

Even if this was achieved, the rabbithole would continue though, because the thing you measure with could now have a backdoor. Remind me of the classic paper about the same problem with software: https://www.cs.cmu.edu/~rdriley/487/papers/Thompson_1984_Ref...