Universal Interpreter: Formalism, Latent Double, and Non-Circular Validation in Combinatorial Spaces

This paper presents an operational framework for the Universal Interpreter (UI)—a procedure that treats every image as a candidate-element of a finite combinatorial space of variations with repetition. Meaning and plausibility of candidates are evaluated through structural descriptors, combinatorial navigation, and neural models (variational autoencoders, VAE) that learn local clusters around real images—anchors. The paper provides explicit definitions, formalizes the local density hypothesis as a testable statistical claim, introduces the principle of limited identifiability of candidates without external reference, and presents an implementation with empirical results. The local density hypothesis is confirmed both on an internal VAE proxy (p < 0.001, Cohen’s d > 4.9) and, methodologically more importantly, on an independent frozen CLIP ViT-B/32 model (Cohen’s d = 1.45–2.18, p < 0.001 after Bonferroni correction). This eliminates evaluation circularity for this test. The latent double experiment demonstrates that the system, with iterative improvement, finds 100% of held-out portraits in VAE space, while blind sampling in CLIP space achieves 95–96/100 hits at threshold τ = 0.78. The key methodological innovation is the introduction of an independent external evaluator (CLIP, with planned extension to DINOv2 and LPIPS) that eliminates circularity. The scope of the protocol is explicitly limited: the UI produces and ranks candidates; verification only establishes class membership, never independent entity existence. Keywords: Universal Interpreter, latent double, combinatorial space, anchoring of real signals, variational autoencoders, identifiability, epistemology of verification, manifold hypothesis, non-circular evaluation.