System Architecture Guide

The Being-in-itself: core question and system architecture

The organism is a 90-dimensional ODE system: 10 metabolic variables (S1-S5, C1-C3, P1-P2) + 80 CTRNN neuron activations, all solved together by an adaptive ODE integrator (Tsit5). The metabolism is a 20-reaction network organized in four layers of chemical structure. The brain reads 41-dimensional sensory input and produces 3-dimensional motor output + 4-dimensional voice output. Motor modulates how the organism interacts with a 300×300 toroidal environment.

System Architecture Diagram

What This Is Not

This is not artificial general intelligence. It is not self-consciousness. It is a functional analog of Hegel's phenomenological progression, implemented as dynamical systems. The scientific question is whether gradient-free, viability-constrained self-maintenance can produce behaviors that structurally resemble consciousness stages. If it fails, documenting why it fails is equally valuable.

The 10-variable, 20-reaction self-maintaining chemical network

The metabolic core is a system of 10 coupled ordinary differential equations with 20 mass-action reactions organized in four layers of chemical structure: Surface (easily learnable), Intermediate (delayed effects), Catalyst repair (cross-catalytic), and Hidden regulator-modulated (not sensed by the organism). Five substrates (S1-S5) are consumed from the environment, three catalysts (C1-C3) enable reactions and repair each other, and two intermediates (P1-P2) mediate delayed effects.

Metabolic Variables

Viability: Life and Death

The viability margin measures how far the organism is from death. It is the minimum margin across all 5 concentrations relative to their death thresholds. A margin of 100% means perfectly safe. A margin of 0% means at the boundary. Negative means dead.

Death is permanent. Once any concentration drops below its threshold, the organism dies and cannot recover. This is a core design principle from viability theory (Aubin): the system must actively maintain itself within a viable region of state space. There is no resurrection.

Perturbations & Toxic Zones

Perturbations target all 10 metabolic variables (S1-S5, C1-C3, P1-P2). Hitting one variable triggers correlated cascades in 1-3 related variables with 300-500 time unit lags, creating causal laws for the organism to discover:

• Catalyst knockout/halving: C1, C2, or C3 suddenly drops. The cross-repair network must compensate.
• Substrate depletion: Any S drops sharply. Motor behavior must adjust resource gathering.
• Noise bursts: Random perturbations across multiple variables simultaneously.
• Input cutoff: Environmental input temporarily stops for a substrate.

At Layer 3+, toxic zones appear: 5-10 wastelands (radius 35-60) and corridors (radius 25-40) where food=0 and enzyme inhibition suppresses catalyst repair. A coupling ramp (coupling_alpha = clamp(t/60000, 0, 1)) gives organisms 60K time steps to learn avoidance before zones become fully lethal. Zones are determinate negation: they must be avoided, not survived through.

CTRNN substrate, Hebbian learning, and predictive coding

The Brain: CTRNN

The brain is a Continuous-Time Recurrent Neural Network (Beer, 1995) with 80 neurons organized in multi-timescale groups (fast τ=1-20, mid τ=20-100, slow τ=100-500). Each neuron follows the dynamics:

The brain receives 41-dimensional sensory input via W_in (41×80) and produces 3-dimensional motor output via W_out (3×80): M1 (intake modulation), M2 (balance/turning), M3 (efficiency). A separate 4-dimensional voice output via W_voice (4×20) provides an expressive social signal.

41-Dimensional Sensory Input

The CTRNN receives a structured 41-dimensional sensory vector combining internal state, temporal dynamics, environment, and social channels:

Hebbian Learning

There is no backpropagation. The CTRNN's weights evolve through four Hebbian rules:

Oja's Rule

Extracts principal components of input. dw = eta * y * (x - y * w). Converges to the first eigenvector of the input correlation matrix.

BCM Theory

Creates selectivity through a sliding threshold. Neurons develop preferences for specific inputs. Below threshold = weakening (LTD), above = strengthening (LTP).

Anti-Hebbian + Competitive

Lateral inhibition decorrelates neuron responses. Different neurons learn to respond to different features.

Neuromodulated Gating

Raw Hebbian learning was found to be net harmful before neuromodulation. Unmodulated learning caused drift and instability. The solution: a neuromodulatory gate controls which neurons learn at each moment:

The self-model signal further amplifies competitive W_out learning, with effective motor gate up to 1.5x.

Predictive Coding Network (PCN)

Per-neuron prediction error dynamics (Rao & Ballard 1999) modulate which neurons learn. This prevents drift in accurate predictors and only allows learning where predictions are failing. The PCN is critical: without it, Hebbian learning destroys rather than helps.

Voice Layer

A 4-dimensional expressive output layer with Oja learning gated by max(stress, self_model_signal). The voice does not carry symbolic content; it is an additional motor channel that co-varies with internal state, giving other agents a richer signal to observe.

Prediction, memory, planning, and concept formation

Self-Model Prediction

The world model has been learning to predict the environment. When it fails in ways correlated with the organism's own actions, a separate self-model emerges. The self-model signal is defined as:

When this diverges from zero, it indicates self-caused effects: the organism's own actions are creating consequences in the world that are distinct from environmental noise.

Self-Model Architecture

The self-model uses random projection (64×83 fixed matrix) to compress the 80-neuron + 3-motor state into a 64-dim representation, then LMS learning (W_read: 41×64) to predict the next sensory input. The self-model prediction is subtracted from sensory input before the world-model prediction step — self-knowledge becomes functional: the organism factors out its own effects when predicting the environment.

A structural signal detects self-causation via correlation: structural_signal = max(corr(residual, motor_dev), max_i(|corr(residual_i, motor_dev)|)) where residual = error not explained by environment (EMA α=0.005) and motor_dev = motor deviation from its own pattern (EMA α=0.02).

Functional Self-Knowledge (Desire)

Self-knowledge becomes functional through determinate negation: when the self-model detects a metabolic deficit, it inverts itself to derive corrective motor output. The organism does not just know what it does — it uses that knowledge to fix what is wrong. correction_benefit measures whether desire_motor actually reduces the deficit.

Episodic Memory

A 128-slot ring buffer storing sensory (41-dim), motor (3-dim), outcome (10-dim), and time for each episode. Cosine-similarity retrieval: when current situation matches a past one (cosine > 0.5 retrieval, > 0.7 influence threshold), past motor actions influence present behavior through motor blending (strength 0.15 × sm_confidence). This is Hegel's Erinnerung — the organism's history informing its present through internalized experience.

Planning

Self-model chaining for single-step lookahead. The organism evaluates candidate actions by predicted viability impact: "If I do X, will my viability improve?" This is satisficing, not optimizing — the organism seeks actions that maintain life, not maximize any fitness function.

Concept Formation

K=8 competitive prototypes over hidden state. Each prototype accumulates motor/outcome statistics through competitive Hebbian clustering. This is Hegel's Verstand (Understanding) — subsuming particulars under universals. The organism groups similar situations and recalls what worked in each category.

Nine stages from metabolism to observing reason

The system progresses through nine layers, each with empirically measurable advancement criteria. Each layer builds on the previous: you cannot have self-knowledge without a brain, you cannot have recognition without self-knowledge.

Encounter, perception, and mutual tracking

The Encounter (Layer 7)

Two organisms are placed in a shared environment with competitive resources. Motor output determines resource capture share, creating natural asymmetry from motor strategy differences. Each organism perceives the 2 nearest others through 14 delayed sensory channels carrying motor (3-dim) and voice (4-dim) output.

Prediction Error Partitioning

Prediction errors are partitioned: other_error tracks error on other-agent channels only. There is no separate "other-model" module — the Hebbian network handles it. The organism's existing prediction machinery encounters something genuinely unpredictable: another autonomous agent.

Recognition Dynamics (Layer 8)

The recognition variable r tracks how much the organism's existence depends on the Other. It is driven by other_pred_error: the organism perceives but cannot predict. A mutual feedback term creates a positive loop: if B recognizes A, A is more likely to recognize B.

Natural selection and structured resource dynamics

Natural Selection (Not Optimization)

Evolution uses natural selection only: binary survive/die. There is no CMA-ES, no tournament, no fitness ranking. This is a philosophical constraint: no external fitness landscape optimization. Organisms that survive pass on their genes. Organisms that die do not.

Autopilot Cycle

The system runs an automated cycle:

• Warmup: 5 organisms to establish baseline
• Evolve: 20-population natural selection
• Run: 25 organisms with evolved parameters
• Re-evolve: repeat from evolved state

Structured Environment (300×300 Torus)

The environment is a 300×300 toroidal grid with 5 independent resource fields (one per substrate S1-S5). Resources regenerate (rate 0.003), diffuse (0.03), and deplete locally. Each field cycles with distinct periods (6K-15K time units) and phase offsets, creating temporal patterns discoverable via Hebbian learning. Cross-resource interaction: eating S_k depletes S_k+2 by 15% — trade-off dynamics. Perturbations come as correlated cascades: when one variable is hit, 1-3 related variables are affected with 300-500 time unit lags.

Empirical findings from the simulation experiments