This paper studies a declaration-first way to choose architecture families. Given a declared universe model, such as an environment, dataset, or dynamical regime with stated observables, the framework asks which information-processing structures are best supported under that declaration. It also frames a longer-horizon inverse question: given a declared structure, which universe classes would support it? The motivating problem is that architectures can match benchmark accuracy while making different commitments about scale, locality, resource use, and failure. Accuracy alone does not decide between those commitments. The proposed procedure treats the unresolved choice as an architectural form of Duhem-Quine underdetermination: the auxiliary commitments under revision are themselves architectural. It fixes the regime, observables, admissible families, resource limits, and evidence tests before asking which family the regime supports. The output is a search-and-narrowing record, not a universal optimizer or polynomial-time algorithm. Within one declared universe model, the argument imports Fisher information geometry as the local measure of statistical distinguishability and Stiefel constraint geometry when inner-product-preserving operators belong to the declared semantics. A conditional transfer from regime statistics to family priority is then gated by written checks for the estimator package, regime evidence, and geometry transport. The paper separates proved results, assumption-bound claims, conditional or reproducibility claims, and open problems. Architecture-family choice is therefore treated as an auditable inference target rather than as a convention inherited downstream.
Piotr Mikołajczyk (2026) studied this question.