Predictive Propagation: A Dual-timescale Architecture for Continuous Environment Flow and Discrete Memory

Overview Core Premise: Predictive Propagation is a dual-timescale cognitive architecture designed to bridge the computational gap between continuous physical motor control and discrete, turn-based sequence models. Hybrid Architecture: It unifies continuous local Spiking Neural Network (SNN) dynamics with discrete, event-driven Large Language Model (LLM) memory consolidation. The Poincaré Phase Boundary: The system uses Umwelt-bounded directional receptor projections and a phase boundary mechanism to collapse continuous prediction errors into a compressed cognitive ledger only when causal surprise thresholds are crossed. Homeostatic & Neuromodulatory Drive: Grounded in Homeostatic Reinforcement Learning, the agent optimizes local deviation relative to a fixed internal setpoint rather than external scalar rewards, utilizing a multi-neuromodulator control framework. Alignment & Stability: Employs scheduled synaptic scaling blackout mechanisms (simulating biological sleep/homeostasis) to resolve catastrophic forgetting and eliminate training-correlated shutdown aversion. .