Problems are often "sticky" (A term I've read referencing this exact thing, a problem where you need to begin implementing before you understand it fully, but I don't think it's a widely used term) but I think the implementation step is a different matter than architecture stuff regardless. It's not that there aren't important things to work out, but for me I broadly consider that to be part of the implementation.
I get what you're saying though, I definitely don't think I'm solving all my problems before ever starting the code, but I definitely always finish solving the problem before I finish the code, and usually early enough that the rest of implementation drags even if there are quirks and edge cases to work out.
Just how much corruption is known. Even if you make the assumption that people in aggregate can generally assess whether corruption exists somewhere, it'd be harder to measure the 'severity' that way. I actually think you're probably correct that people's perceptions of corruption have some connection to reality, but I think the more fundamental problem with such a scale is boiling a complicated reality like corruption into a single corruption number, which seems pretty silly on its face.
I read something interesting recently about how most discussion of brain function is 'tainted' by metaphor, these days usually involving computers. Basically the concept was that our brains don't process and store information like a computer does, but because computers come the closest to accomplishing what our brains do it's the easiest point of reference for us when discussing brain function.
Since I don't really know anything about brain science and know a little about computer science I find it pretty hard to escape that metaphor and imagine how a brain can function without doing what a computer does. This type of article does give me a bit of an inkling though.
I get what you're saying though, I definitely don't think I'm solving all my problems before ever starting the code, but I definitely always finish solving the problem before I finish the code, and usually early enough that the rest of implementation drags even if there are quirks and edge cases to work out.