It doesn’t seem to me like your prior needs to encode information about the “true integers” any more than it needs to encode information about the world. It just needs to be flexible enough to see the structure that is there.
Maybe this is what you mean by “if we just need a prior,” but it seems like an important point.
While we’re at it, it’s not even clear our prior needs to capture PA. Maybe you could use something much weaker, perhaps even a theory so weak it could assert its own (cut free) consistency, and then just learn the structure of PA by observing some computations (or however you gain info about logical facts).
It doesn’t seem to me like your prior needs to encode information about the “true integers” any more than it needs to encode information about the world. It just needs to be flexible enough to see the structure that is there.
Maybe this is what you mean by “if we just need a prior,” but it seems like an important point.
While we’re at it, it’s not even clear our prior needs to capture PA. Maybe you could use something much weaker, perhaps even a theory so weak it could assert its own (cut free) consistency, and then just learn the structure of PA by observing some computations (or however you gain info about logical facts).