I thought a bit about the question https://mathoverflow.net/q/503195/ and arrived at the following little problem of computability. Any ideas? It happens to involve something like the “parameterized realizability” from nprofile1qyt8wumn8ghj7un9d3shjtnyd968gmewwp6kytcqyz4g9497lls2ey4v3rxtld6a8m5rh7meav97tqfhnceyja6wn0lpgng99up (nprofile…99up)'s and nprofile1qyt8wumn8ghj7un9d3shjtnyd968gmewwp6kytcqyq0gte8xzp92t6508mavz7g8z37mv2c2n6sgfw5luws2j8m7nr7ez64kgh3 (nprofile…kgh3)'s paper https://arxiv.org/abs/2404.01256.
Let A, B ⊆ ℕ be disjoint. Assume that there is a program M which, given x ∈ A ∪ B, decides whether x ∈ A or x ∈ B, but with some help. Namely, there is a “desired oracle” which is an infinite word on four letters, and M has access to a “noisy” version of this oracle, which can be anything satisfying a requirement (that's the “parameterized” part: M must work given any oracle from a certain set; but unlike the countable reals paper, here the set will also depend on the input x). The requirement is the following: Arrange the four letters into a 2×2 table. Then if x ∈ A, the n-th letters of the desired and actual oracle must be on the same line of the table, while if x ∈ B they must be on the same column.
The question is: Must there be another program M' which does the same but without the help?
The difficulty is this: If M “knows” the n-th letter of the desired oracle (it manages to compute it somehow), and the n-th letter of the actual oracle is different, then M can deduce whether x ∈ A or x ∈ B. But if the actual oracle is the same as the desired one, then M can use that as computational power.
If I try to create M', I'd like to make it run M on well-chosen actual oracles. Those letters of the desired oracle which M “knows” should match in the actual oracle. But M' can't make everything match because it also has to work computably. So how to extract what exactly M knows in order to replicate it?