Project MA: A Deterministic, Energy-Constrained Framework for Probabilistic Cognitive Modeling

Project MA presents a deterministic, energy-constrained cognitive architecture for studying self-organizing intelligence under local metabolic and entropic constraints. This work introduces a framework for investigating cognitive dynamics through agent-based modeling, departing fundamentally from contemporary artificial neural networks trained via global loss minimization and backpropagation. Instead, cognition is modeled as an emergent property of large populations of autonomous, homeostatic computational units (“MA-Neurons”) governed exclusively by local rules. Each MA-Neuron operates as an independent agent that regulates its internal metabolic energy budget, estimates informational uncertainty via the Shannon entropy of its recent activation history, and adapts structurally through plasticity mechanisms combining Hebbian potentiation with heterosynaptic depression. Network topology is not fixed but treated as a first-class state variable that continuously evolves in response to signal novelty, enabling ongoing synaptic growth and pruning without catastrophic forgetting. Learning arises purely from local interactions; no global objective function, centralized optimizer, or gradient-based learning rule is employed. The architecture synthesizes principles from neuro-inspired computation, agent-based modeling, and sparse graph dynamics. It is implemented as a high-performance, event-driven simulation with strict determinism enforced through seeded pseudo-randomness, enabling bitwise-reproducible execution across platforms. Despite asynchronous signal propagation, causal ordering is preserved via a centralized priority queue, allowing complete replay and inspection of complex cognitive trajectories. Wake and dream phases are explicitly modeled to separate constrained sensory integration from unconstrained consolidation and structural stabilization, supporting fine-grained causal analysis and addressing key challenges of explainability and reproducibility in complex adaptive systems. Empirical results documented in this version (v1.0) demonstrate several verifiable emergent behaviors derived solely from local interactions. These include (i) stable minimization of system-level entropy under metabolic constraints, (ii) a pronounced non-linear resonance phenomenon observed at intermediate network scales (approximately N ≈ 500), and (iii) long-term structural stability arising from balanced plasticity dynamics. These behaviors emerge without task-specific training or external supervision, suggesting that cognitively relevant dynamics can arise from minimal survival-driven principles alone. This preprint is released as an open, reproducible scientific artifact intended to serve as a rigorous testbed for theoretical research in computational neuroscience, cognitive architectures, active inference–inspired systems, and non-gradient-based learning paradigms. While the architecture draws inspiration from biological principles, it makes no claims regarding consciousness, sentience, or artificial general intelligence; all cognitive terminology is used strictly as a mathematical and operational abstraction. All source code, experimental datasets (CSV), and plotting scripts associated with this work are publicly archived and versioned on Zenodo to facilitate independent replication, parameter ablation, and extension by the research community. Code repository: Github Archived release (paper, code snapshot, datasets): (archived release)