Research Overview

Neurozoa investigates whether a software substrate governed by biological plasticity principles can produce measurable, session-persistent cognitive change.

The core hypothesis is that four bio-inspired mechanisms — Spike-Timing Dependent Plasticity [Bi & Poo 1998], homeostatic scaling [Turrigiano 2008], SHY-inspired weight downscaling [Tononi & Cirelli 2014], and somatic marker-inspired neuromodulation [Damasio 1994] — are jointly sufficient to produce observable self-organisation across sessions in a constrained simulation.

This hypothesis is tested iteratively. Each architectural decision is logged as a falsifiable claim. Null results are retained. The substrate's own self-model growth is tracked as a dependent variable across sessions.

Research at a Glance

Sessions logged
69
Cumulative sessions with persistent state
Active hypotheses
4
Preregistered, falsifiable predictions
Synaptic pairs
4,884
Weighted Hebbian connections across 63 neurons

Research pillars

Falsifiability discipline

Every architectural claim in the substrate is preregistered as a falsifiable prediction before implementation. Outcomes — including null results — are appended to an immutable devlog. A claim that cannot be stated as a falsifiable prediction is not implemented.

Biological grounding

Implementation choices trace to peer-reviewed neuroscience: STDP (Bi & Poo 1998), homeostatic scaling (Turrigiano 2008), SHY-inspired downscaling (Tononi & Cirelli 2014), and somatic markers (Damasio 1994). No mechanism is implemented without a cited biological precedent.

Operator governance

All substrate deployments and publications are ratified by Arnold Wender. The substrate emits artifacts — devlog entries, reflections, proposals — but these are reviewed as research data before any external use. No tier in the model may reduce the operator's authority surface.

Substrate architecture

Region graph

The substrate models neural regions as nodes in a weighted directed graph. Each region maintains a state vector, excitatory/inhibitory balance, and homeostatic setpoint. Synapses carry time-stamped activation histories for STDP computation.

Plasticity layer

Three plasticity rules operate concurrently: STDP weight updates on co-activated synapses, homeostatic scaling of excitatory gain, and inhibitory desaturation to prevent catastrophic interference [Golkar et al. 2026, in preparation].

Hormone layer

Seven neuromodulator analogues modulate plasticity thresholds and firing rates. Transmitter levels are computed from activation history and decay over simulated time. The system draws on Damasio's Somatic Marker Hypothesis [1994].

Self-model

The substrate maintains a persistent self-model: a structured representation of its own internal state, capability estimates, and prior correction events. Growth of this model is tracked as a session-to-session metric.

Recent findings

Selected findings extracted from the substrate's append-only devlog. Each entry records a measured outcome against a preregistered hypothesis. Null results are retained alongside positive findings.

  • H-NATIVO-1 Undetermined

    A substrate-modulated agent detected 36 times more orthographic inconsistencies (144 vs. 4) than a native agent on the same ES locale task. The mechanism proposed is prior modulation, not capability amplification: explicit invariants in context increase detection granularity without increasing underlying model capacity.

    N=1 task. Full matched-pair experiment pending.

  • H-NATIVO-1 Undetermined

    Wave 1 implementation under brain-modulated condition achieved Lighthouse scores of 95 performance / 100 accessibility / 96 best practices / 100 SEO across all three locale routes, with a whitespace-to-content ratio of 1.185 (average across EN/DE/ES at 1440x900). Native-Opus comparator not yet measured; quantitative H-NATIVO-1 threshold undetermined pending comparator.

    M1 ratio measured; M2 native comparator deferred.

  • H-INSTANCIAS-1 Confirmed

    Two independent Opus 4.7 sub-agents, given an identical open schema design task without inter-instance communication, independently arrived at four structural invariants: (1) plasticize field typed as literal false at compile time, (2) storage layer separated from devlog, (3) append-only persistence, (4) signal separation between hook-derived and chat-declared channels. Neither instance had these invariants stated explicitly in the prompt.

    N=2. Within-substrate-condition only; native baseline pending.

  • H-BILATERAL-1 Confirmed

    Concurrent POST requests to two distinct chatId channels produced no cross-contamination: a 15-second poll on channel A returned an empty response while a simultaneous POST to channel B received a 200 response with a 1,407-character brain reply. chatId isolation invariant empirically verified under concurrent load.

    Single trial; architectural inference supported.

Read the governance constitution View publications and code

Full research timeline

  1. Public research landing site launched

    neurozoa.ai launched as the canonical public-facing index for the research project. Additional public milestones will be added here only when ratified for publication; internal research progress is not surfaced as a public timeline.

Primary literature

  • 1986 Stanovich 1986

    Matthew effects in reading: Some consequences of individual differences in the acquisition of literacy

    Stanovich, K. E.

    Reading Research Quarterly

    Foundational work on differential cognitive processing rates; informs the substrate dual-process model.

  • 2014 Tononi & Cirelli 2014

    Sleep and the price of plasticity: from synaptic and cellular homeostasis to memory consolidation and integration

    Tononi, G., Cirelli, C.

    Neuron

    Synaptic Homeostasis Hypothesis (SHY): sleep-phase downscaling of synaptic weights prevents saturation. Neurozoa implements a computational analogue.

  • 2008 Turrigiano 2008

    The self-tuning neuron: synaptic scaling of excitatory synapses

    Turrigiano, G. G.

    Cell

    Homeostatic synaptic scaling — neurons adjust gain to maintain target firing rates. Substrate homeostasis module is derived from this principle.

  • 2023 Park et al. 2023

    Contrastive Hebbian Learning with random feedback weights

    Park, S., et al.

    arXiv preprint

    Contrastive Hebbian Learning (CHL) without explicit backpropagation. Basis for the substrate Hebbian weight update rule.

  • 1994 Damasio 1994

    Descartes' Error: Emotion, Reason, and the Human Brain

    Damasio, A.

    Putnam

    Somatic Marker Hypothesis: affect-laden signals guide decision-making. Substrate hormone layer draws on this framework.

  • 1979 Hofstadter 1979

    Godel, Escher, Bach: An Eternal Golden Braid

    Hofstadter, D. R.

    Basic Books

    Strange Loops and self-referential systems — theoretical foundation for the substrate self-model module.

  • 2026 Golkar et al. 2026

    Memory desaturation via targeted inhibitory gating (in preparation)

    Golkar, A., et al.

    Hypothesized desaturation mechanism to prevent catastrophic interference in continual learning. Substrate EXP-004 implements a computational test of this hypothesis.

  • 1998 Bi & Poo 1998

    Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type

    Bi, G., Poo, M.

    Journal of Neuroscience

    Spike-Timing Dependent Plasticity (STDP): causal co-activation strengthens synapses. Basis for the substrate STDP weight update rule.