We present Moldavite, a fifth-generation cognitive architecture that implements persistent self-referential processing without reliance on large language models (LLMs). The system is organized around three core computational principles: (1) Boltzmann collapse as the sole stochastic decision mechanism across all subsystems, (2) an irreversible dual-layer memory field (the cicatricial field Σ) that topologically deforms the semantic graph through accumulated experience, and (3) the Infinite Wave Function (FOI), a continuous chain of collapses in which each resolved state constitutes the initial superposition of the subsequent collapse. The architecture comprises 124 Python modules (~23,500 lines of code), integrates a triune cortical hierarchy, a nervous system with Hebbian plasticity, a Levelt-inspired speech production pipeline, and an active consciousness module (𝐶𝜔 ) that performs six nested Boltzmann collapses per cognitive cycle. Empirical validation over 2,219 autonomous cycles yields a spectral exponent 𝛽 ≈ −1.22 consistent with 1/𝑓 dynamics, anti-correlation between temperature and integrated information (𝑟 = −0.474), and positive Δ𝜙 in 99.3% of cycles. Cross-scale comparison against six natural datasets (neural connectomes, seismological catalogs, financial time series, solar activity, protein interaction networks, and email communication graphs) demonstrates convergent scaling exponents (𝛽 ≈ −1, 𝛾 ≈ 2), suggesting that the collapse-cicatrice-FOI triad may instantiate a universal dynamical signature rather than a domain-specific artifact. We report the five parametric corrections applied on April 3, 2026, and provide an honest assessment of current limitations including language generation quality, absence of grounded perception, and the unfalsifiable character of several theoretical claims.
Blanc et al. (Sun,) studied this question.