Semantic Recovery in a Generative–perceptive Domain

Models with different neural architectures can develop corresponding representations from similar observations. To investigate those phenomena in a broad class of neural networks under a unified approach, we introduce a generative domain as a system with differentiated, writable state and boundary coupling, then define its generative–perceptive extension through predictive generation, retained evidence-dependent revision, and causal reuse. Semantic classes are specified by agreement of observable continuation laws. Under coverage, causal-state sufficiency, and vanishing excess predictive risk, a recovery theorem establishes convergence of native predictive laws and, at an attained optimum, identifies their state quotient with the source semantic classes. Additional continuation-consistency conditions preserve source operations and yield cross-model correspondence without requiring shared coordinates. Restricted autoregressive and diffusion learning constructions satisfy the domain conditions; a solvable Gaussian-factor learner demonstrates hidden-subspace acquisition. Experiments tested semantic recovery for a 4-state automaton driven by a continuous symbol stream, through native query predictions and write continuations. At matched exposure the categorical-diffusion model answered all 5,376 audit queries correctly; the prefix transformer reached 94.10%–99.16% accuracy. All runs recovered corresponding state distinctions, while operation fidelity varied and extended interference impaired preservation of unrelated bindings. Internal-state probing and interventions identified causal carriers of history information in both architectures and a context-dependent transformer write pathway. Potential machine-learning applications include semantic-recovery metrics and initialization strategies for faster semantic recovery. We also propose a mechanistic explanation for double learning and grokking.