Mathematical Proof Beyond Symbolic Proof

Both formalism and Platonism hold that the symbolic expression of a proof is either the proof itself or an adequate vehicle for it. I argue that both are wrong in the same way, but that the correct response is epistemic rather than ontological. Drawing on the representational geometry of large-scale neural networks, I show that a single mathematical truth can be accessed through two architecturally distinct modes--symbolic serialisation and high-dimensional tensor geometry--whose correspondence is verifiable and whose internal procedures are mutually unavailable. The unavailability is not merely practical in the sense of awaiting better tools. The tokenisation bottleneck establishes an epistemic asymmetry between representational architectures that cannot be closed under feasible verification constraints--not a translation problem at the level of practicable mathematical agency. The position this establishes is constraint realism with architectural epistemic pluralism. Mathematical reality is real in a specific and limited sense: it is genuinely constraining, forcing representations into correspondence regardless of the architecture in which they are made. But no single architecture exhausts access to the constraint space. Different architectures have different access profiles over the same mathematical reality--partial, mutually irreducible, and bounded in ways that are geometric rather than merely practical. This is the sense in which the argument echoes, at the level of representational architectures, what Gödel’s incompleteness theorems establish at the level of formal systems. The cost of accepting this argument is the loss of what might be called architectural parochialism--the assumption that the boundary of symbolic mathematics coincides with the boundary of mathematical reality itself. The map is not the territory. In mathematics, as elsewhere, the territory is larger than any single map reveals.