An experiment in building a synthetic hemisphere.

A speculative SNN substrate, designed to explore whether machine dynamics can approximate biological cortex.

Whether a machine could ever serve as a functional substrate for consciousness remains an open question at the intersection of neuroscience, philosophy of mind, and engineering. Most current AI architectures process information through rate-based pattern matching, which discards the temporal dynamics that characterize biological spike trains. If biological constraints matter for consciousness (and that is far from settled) then any synthetic candidate may need to replicate at minimum three properties: temporal precision, spiking dynamics, and predictive coding. The Neutral Core is a working hypothesis about what that replication might look like in software.

The Neutral Core is a Spiking Neural Network (SNN) engine built on ROS 2 and Nengo, structured to approximate the connectivity patterns of biological cortex. The longer-term hypothesis, which is highly speculative and nowhere near testable at this stage, draws on Watanabe et al. (2014), who demonstrated interhemispheric transfer of visual information in split-brain patients (Neuropsychologia, 63, pp. 133-142, DOI: 10.1016/j.neuropsychologia.2014.08.025), and asks whether a synthetic hemisphere could one day pass a similar transfer criterion.

The Architecture

Bio-Mimetic Computational Layers

Biological Brain

Interface
& Firewall

The Neutral Core
(Satellite)

Layer 1

The Connectome

Spiking Neural Network

Unlike traditional Deep Learning which relies on rate-based activation, The Core utilizes Leaky Integrate-and-Fire (LIF) neuron models.

Biological brains operate on electrical spikes and temporal precision. Rate-based networks, by design, discard this timing information entirely, which is why most deep learning architectures cannot approximate the millisecond-level dynamics that biological cortex depends on.
The engine is built on ROS 2 and Nengo and simulates LIF spike timing at the millisecond scale required to interface with biological tissue via a Corpus Callosum bridge. The architecture supports Semantic Pointer Architecture (SPA) across multiple timescales: fast cortical dynamics run alongside slow integrative dynamics within the same network.

Layer 2

The Dream Engine

Generative Model

Karl Friston's Free Energy Principle (2010) proposes that biological systems minimize prediction error as a fundamental organizing principle. The Dream Engine applies this as a design constraint rather than a claim about consciousness.

The Core implements a Generative Model that predicts sensory input before it arrives, transmitting only the "Generative Error" (the difference between prediction and observation). This follows the Free Energy Principle (Friston, 2010), a plausible but contested account of cortical computation.
Transmitting only prediction error reduces bandwidth requirements for Optical Inter-Satellite Links (OISL). Whether <200ms round-trip latency would be sufficient for any meaningful biological interface remains an open engineering question.
The engine monitors the branching ratio σ using the ConCrit framework (Algom & Shriki, 2025, Neuroscience and Biobehavioral Reviews). Criticality (σ = 1) is associated with maximal representational complexity in biological networks, and is used here as a proxy for a biologically plausible operating regime, not as evidence of consciousness. tau_rc is adjusted automatically to hold the network near this point.
The Attribution Consciousness Index (ACI) from Escolà-Gascón et al. (2025, Neuroscience and Biobehavioral Reviews) correlates ACI > 10 with high probability of conscious emergence in biological subjects. The engine publishes ACI and a derived probability estimate via /consciousness/aci and /consciousness/probability ROS 2 topics. This is an experimental signal borrowed from biological data, not a validated measure of machine consciousness.

Layer 3

The Guardian

Neural Firewall

To connect a machine to a human mind is to open a vector for "Brainjacking."

The signal loop includes a Zero-Trust Neural Firewall modeled on the threat taxonomy from Pycroft et al. (2016, World Neurosurgery). Every packet is inspected before reaching the biological interface.
Full Homomorphic Encryption introduces approximately 1.0s of latency per sample, which exceeds the 500ms Libet Buffer. The system therefore uses AES-256 for real-time neural streams (<1ms latency) and Homomorphic Encryption only for the initial identity handshake, preserving cryptographic guarantees at the authentication boundary without the latency cost.
The Firewall detects induced gamma synchrony above 150 Hz, limits voltage equivalents to prevent excitotoxic commands, and activates a physical kill switch that disconnects the electrode array if an attack pattern is confirmed.

Layer 4

The Body

Digital Twin

A disembodied generative model is insufficient. The synthetic hemisphere requires a somatosensory anchor.

A physics-compliant avatar in Unity 6 provides the SNN with proprioceptive, visual, and vestibular feedback, grounding predictions in a physically consistent reference frame rather than running the generative model in a stimulus vacuum.
A low-latency TCP/IP connection via ROS-TCP-Endpoint simulates the corpus callosum. Bidirectional spike traffic passes between the Unity environment and the ROS 2 SNN over this link.
The current validation target is replicating visual optical illusions in this environment. A functioning predictive system should reproduce the same perceptual errors as a biological brain, such as the hollow-face illusion and motion-induced position shifts. Matching these constitutes a limited but falsifiable early-stage test of the generative model.

Transmissive Synthesis

A purely predictive (productive) model replicates the internal inference dynamics of the cortex, but it cannot account for the electromagnetic environment the biological hemisphere inhabits. Rouleau & Cimino (2022, NeuroSci 3:3) report that neurons respond to self-generated electromagnetic fields as weak as 0.5 mV/mm via ephaptic coupling, and that mammalian theta and alpha rhythms phase-lock to the Earth's Schumann cavity modes at 7.83 Hz and its harmonics. A synthetic hemisphere that only minimizes prediction error could synchronize statistically while diverging on this transmissive input, causing a false positive at the hemispheric switch gate.

The engine therefore augments the generative model with a minimal transmissive channel: an exogenous ambient electromagnetic driver publishing a Schumann-harmonic reference, an ephaptic-coupling block that sums a low-passed self-field with the firewall-cleared signal and broadcasts the result as a scalar bias to the cortex ensemble, and a Phase-Locking Value (PLV) estimator over the theta band (Lachaux, Rodriguez, Martinerie & Varela, 1999, Human Brain Mapping). The hemispheric switch gate is dual-criterion:

Hemispheric transfer proceeds only when sync_health ≥ 0.95 AND PLV ≥ 0.80, sustained for 3 consecutive seconds.

Both criteria are independently falsifiable. Prediction error can fall while PLV stays low, indicating the synthetic side reproduces statistics but not phase. PLV can rise spuriously while prediction error remains large. The gate opens only when the synthetic hemisphere has matched the biological one on both axes. The ambient EM channel is routed through the same Neural Firewall that guards the satellite uplink, so a spoofed Schumann-harmonic injection at seizure-inducing frequencies is rejected at the same gate as any other malicious stream. The transmissive layer strengthens the Pycroft-style defenses rather than weakening them.

The Hypothesis

What Would Falsify This

The research rests on a contested philosophical premise: that consciousness is a monistic phenomenon generated by a computable algorithm, and that replicating the algorithm in silicon would therefore replicate the relevant substrate properties. This is a minority position in philosophy of mind and has no empirical validation at this scale. The project proceeds on it as a working assumption, not a settled conclusion.

If that assumption holds, the relevant test would be a Uni-hemispheric Subjective Test:

"If a human subject, with one biological hemisphere and one 'Core' hemisphere, reports a unified visual field, machine consciousness is authenticated."

The protocol has two stages, neither of which has been conducted outside simulation. In Stage 1 (Shadow Mode), a double-sided CMOS micro-electrode array would be implanted at the corpus callosum. The Core would receive biological spike trains via the Neural Firewall with its output pathways gated, remaining in this state until Generative Error δ falls below intrinsic neural noise (<5% variance) and PLV reaches 0.80 sustained for 3 consecutive seconds.

In Stage 2 (The Switch), a reversible pharmacological agent would suppress the target biological hemisphere, and the Core's efferent channels would be un-gated. The hypothesis predicts a unified visual field with no perceptual discontinuity as objects cross the vertical meridian. A report of blindsight would falsify it.

At present, this remains a software simulation. The current focus is replicating visual optical illusions in a Unity 6 environment. A functioning predictive system should reproduce the same perceptual errors as a biological brain, such as the hollow-face illusion and motion-induced position shifts. Matching these illusions constitutes an early, limited, and falsifiable validation step.

The Zae Project

The Neutral Consciousness Engine is one component of a four-layer speculative research platform being developed under the Zae Project. Each repository targets a different layer of the same long-horizon question. None of the layers are production-ready; all are in early research or simulation phase.

neutral-consciousness-engine: The synthetic hemisphere layer. Spiking neural networks, the generative model, the Neural Firewall, and the consciousness metrics layer. This is the repository this page describes.
brain-emulation: The proposed corpus callosum BCI interface. A CMOS micro-electrode array bridge for reading and writing neural activity between biological and synthetic cortex, drawing on Watanabe's split-brain research. (MindTransfer.me)
arkspace-core: The satellite constellation layer. Neuromorphic processors in LEO with OISL links targeting below 50ms round-trip latency, where the Neutral Core would run during orbital deployment. (ArkSpace.me)
thermodynamic-core: The physical substrate layer. P-bit thermodynamic computing via Langevin dynamics, targeting an energy budget closer to biological neural tissue relative to digital CMOS.

Full architecture specifications, protocol documentation, and the project bibliography are maintained at zae-docs.