Biologically-Constrained Modular Active Inference: A Functional Architecture from Trace to Gating (RDLNN v1.0)

This preprint presents the Recurrent Deep Liquid Neural Network (RDLNN) — a fully biologically-constrained, ten-module cognitive architecture engineered for direct deployment on neuromorphic hardware. RDLNN solves the systemic credit-assignment problem without any form of global backpropagation by replacing it with a macroscopic, precision-weighted Variational Free Energy (VFE) scalar computed in the Cingulate Active Inference Hub. This single scalar broadcast (η) structurally gates associative consolidation across all local multi-factor modules, isolating physically meaningful geometric plasticity from pervasive continuous-time background noise. Key Empirical Results • In closed-loop tri-module gating tests, the VFE mechanism produces near-perfect separation (CosSim = 0.9535 in high-surprise vs. 0.0744 in zero-surprise and −0.0651 when η is clamped), with ∆ > 0.90 across noise regimes. • When scaled to the full simultaneous ten-module continuous execution (1 ms ODE integration), the same mechanism widens the separation margin to ∆0.9475 (C1 = 0.9636 vs. C2 baseline = 0.0161), demonstrating robustness rather than degradation under aggregate noise. • Three hard architectural integration invariants are formalised: (1) temporal-discontinuity-only Prediction Error (|x(t) − x(t−1)|), (2) default-open Thalamic bootstrapping for survival-critical heuristics, and (3) numerical stability bounds (η_CB ≤ 1/N_Delay and ∆W ≤ 10⁻³ per step for Amygdala/Cerebellar loops). The architecture strictly adheres to three-factor eligibility traces (τ_e = 500 ms), local β-topology parameters, Kalman-precision stream normalisation, and subcortical reflex circuits (12 ms Amygdala, 41 ms Insula leaky-integrator chains), all validated in continuous-time differential simulation. Relation to Prior Work This standalone functional implementation extends and operationalises the NeuroForge unified neural substrate framework previously released as: https://doi.org/10.5281/zenodo.17622945 (NeuroForge: A Unified Neural Substrate for Scalable Biological AI – Three Preprints + 2-Million-Neuron Technical Report). It provides the first complete, reproducible, hardware-ready specification of one of the core modular architectures outlined in that collection. Keywords: active inference, RDLNN, variational free energy, credit assignment, neuromorphic hardware, three-factor learning, cingulate gating, thalamic bootstrap, continuous-time simulation, biological plausibility, liquid neural networks Additional Notes for Zenodo: • Preprint (March 2026) – fully self-contained with methodology, closed-loop experiments, motor & affective sub-routine validation, and explicit integration invariants. • All figures (Kalman precision tracking, η broadcast, Cosine similarity bars, Basal Ganglia/Cerebellar trajectories, Insula multi-frequency synthesis) are included in the PDF. • Designed for immediate neuromorphic porting; code and simulation parameters will be released in a follow-up repository. This record advances biologically faithful scaling of Active Inference beyond toy models, demonstrating that a single macroscopic VFE scalar is causally sufficient for systemic, noise-robust credit assignment in a fully recurrent, multi-module substrate.