This work identifies a phase transition in latent generative graph models governed by the kurtosis of the radial latent distribution. We show that structural identifiability does not emerge from uniqueness of solutions, but from the appearance of a dominant configuration within a degenerate solution manifold. Across a wide range of latent distributions, the effective degeneracy collapses according to a universal sigmoidal law as a function of kurtosis, with a critical threshold κc ≈ 3. 0. This result demonstrates that identifiability is governed primarily by statistical properties of the latent space, rather than by inference parameters. All source code, data, and figures used in this work are publicly available at: https: //github. com/eduwardus/emptiness-and-identifiability This work was originally motivated by conceptual questions related to Madhyamaka philosophy, particularly the analysis of reification and the emergence of apparent identity from relational structure. The formal results presented here provide a quantitative framework that may serve as a bridge between statistical inference and relational ontologies.
Eduardo Gonzalez-Granda Fernandez (Fri,) studied this question.