Three-Level Hierarchy of Cognitive Emergence

This repository contains the full research work and experimental implementation for a proposed three-level cognitive architecture that models the emergence of structured decision-making in artificial agents. The framework introduces: Level-1 – Reinforcement Layer:A direct reward–signal system that forms the primitive learning substrate. Level-2 – Inefficiency-Aware Multi-Path Consequence Reasoning (IMCR):A prediction-based mechanism that evaluates candidate reasoning paths by forecasting multi-dimensional consequences, enabling agents to reason beyond reward maximization. IMCR operates through a learned consequence predictor and a structural-value evaluator, allowing the system to prioritize stability, coherence, and long-horizon outcomes. Level-3 – Preference Stabilization Dynamics (PSD):A higher-order layer that clusters historically observed consequence patterns, identifies recurring attractor states, computes stability scores, and biases future reasoning toward internally consistent behavior. PSD provides a mechanism for early forms of preference formation, consistency maintenance, and proto-identity shaping. The repository includes a Python implementation that replicates the experiments described in the paper, including: Consequence prediction learning curves, PSD-driven clustering dynamics, Before/after path-selection distributions, Perturbation–recovery behavior, Multi-agent divergence under identical training conditions. These components collectively demonstrate how coherence, internal structure, and early preference-like tendencies can emerge from repeated multi-path reasoning episodes, even in the absence of explicit objective functions.