The Cosmic Information Theory (CIT/TIC) establishes that the spectral structure of the Riemann zeta function governs the thermodynamic stability of the physical information substrate, with the Generalized Riemann Hypothesis emerging as a necessary condition for unitary information conservation. In this work, we propose the Temporium Attention Hypothesis (TAH): the spectral confinement mechanism—prohibition of modes below the mass gap ∆ by the zeta-zero spectrum—constitutes a universal principle of information processing, operating in both physical (quantum, cosmological) and artificial (deep neural) substrates. We computationally validate TAH through a neural attention architecture incorporating three geometric-spectral invariants of CIT/TIC: (i) the pseudo-Riemannian metric gij = ∂i∂j (− log |ζ(s)|) as a similarity function; (ii) the phase navigation ratio θ = t1/t∗ ≈ 2.5416 as a fundamental coherent transport step; and (iii) the spectral gap ∆ as a dynamic sparsity regularizer. On a toy dataset of sequences with “zeta-critical” versus “non-critical” patterns, Temporium Attention achieves: • 67.0% active sparsity (vs. 2.0% for standard softmax), reducing effective connections by 33×; • Noise robustness superiority: 77.5% accuracy under 50% token corruption (vs. 72.0% baseline); • Convergence of CIT parameters: θ → 2.4632 (−3.09% from theoretical 2.5416) and ∆ → 0.02539 (64× scale adaptation), demonstrating that TIC invariants are natural learning attractors. We extend the analysis to propose four application domains of TAH: (1) computational efficiency in large language models; (2) structural interpretability via the E8 lattice; (3) adversarial robustness in vision systems; and (4) hybrid quantum architectures. We also discuss five open problems, including the conjecture that TAH implies fundamental energyefficiency limits for any information processor—physical or artificial—operating below its characteristic spectral gap.
Leandro de Oliveira (2026) studied this question.