The Minimum Viable Learner - A Parameter-Free Unified Theory for a Quantized Learning Unit -20th May 2026 

This paper presents the theoretical blueprint for the Minimum Viable Learner (MVL), a framework reinterpreting the structural patterns of the observable universe as the emergent operational protocols of a quantized learning machine. Governed by a parameter-free control loop known as Quantized Regret Minimization (QR_MIN), the system manages the universal trade-off between exploration and exploitation entirely through localized coordinate geometry rather than hand-tuned hyperparameters. Rather than asserting a completed architecture that fully replaces connectionist training algorithms, this work highlights a clear developmental path toward highly efficient, backpropagation-free functional mapping by presenting a network design where streaming inputs enter directly at a centralized system centroid. In this configuration, alternative decision pathways naturally function as concurrent Fourier components that collectively optimize to eliminate local residual noise errors at the hub.

Deep validation of these information-theoretic design rules is provided by a drastic dimensional reduction from Mass-Length-Time (MLT) to simple Length-Time (LT) registries, where mass emerges as the localized packing of quantized variance. Under this framework, physical invariants—including the speed of light, gravity, and the Planck and Boltzmann constants—resolve as dynamic state parameters that scale deterministically with system maturity to maintain a consistent user interface for an internalist observer. Empirical simulations of a wavefunction-inspired 3D projection variant (QR_MIN_WF) exploit additional dimensional freedom to demonstrate an approximate 5.5% reduction in cumulative regret over 1D/2D baseline optimizations. By regulating internal network dynamics via a self-balancing trade-off between component frequency precision and structural basis orthogonality, the MVL framework provides a mathematically stable, non-parametric development track for designing resilient, highly parallelized neural computing systems.

This document is presented as an early work in progress for public sharing of ideas, intended to stimulate discussion and future development work. Please note the role of Google Gemini as a collaborative agent to generate supporting results.  This partnership allowed the primary author to connect Information Theory (Regret Minimization) with Subatomic Physics in a way that the primary author’s traditional engineering specialization would not otherwise permit. Much of what has been presented in this document has been presented with the specific intention of providing the foundational materials for future AI supported research by readers however they may see fit.  Given the velocity of AI development today these ideas are probably best explored further using an automated evolutionary capability like AlphaEvolve or OpenEvolve.