Encephagen: Research Development Report

Date: April 1-2, 2026 Author: edvatar (toroleapinc) Repo: https://github.com/toroleapinc/encephagen AI-assisted: Code development assisted by Claude (Anthropic). All experimental design, analysis, and interpretation by the author. All code reviewed, validated, and tested.

1. Project Overview

Encephagen investigates whether human brain macro-connectome topology alone — without region-specific parameters — produces emergent functional organization and supports biologically plausible learning.

Starting from real structural connectivity data (Human Connectome Project, 96 regions including subcortical), we built a spiking neural network brain and progressively tested: does topology create functional differentiation? Can the brain learn? Can it control a body?

2. Prior Work (Must Cite in Any Paper)

Paper	Year	What they showed	How we relate
Gollo et al.	2015	Rich-club nodes develop slower dynamics from identical Stuart-Landau oscillators on primate connectome	We extend with two-level decomposition + spiking neurons
Zamora-Lopez & Gilson	2025	Wilson-Cowan with identical params on human connectome shows regional variance/timescale diversity (J Neuroscience)	Most direct predecessor — published March 2025. We add degree vs wiring decomposition
Murray et al.	2014	Empirical timescale hierarchy across primate cortex (Nature Neuroscience)	Our spiking model reproduces this without heterogeneous parameters
Chaudhuri et al.	2015	Timescale hierarchy requires heterogeneous parameter gradient (Neuron)	We show topology alone produces a hierarchy (though less precise)
Cabral et al.	2011	FC from identical Kuramoto oscillators on human connectome (NeuroImage)	We add degree/wiring decomposition
Deco et al.	2014	Optimal SC-FC at bifurcation, homogeneous parameters (J Neuroscience)	We go beyond FC matching to functional identity
Honey et al.	2007	Neural mass models on macaque connectome vs rewired surrogates (PNAS)	Foundational — our work extends to spiking neurons
Váša & Mišić	2022	Null models in network neuroscience (Nature Reviews Neuroscience)	Methodological framework for our null model approach
Lappalainen et al.	2024	Drosophila connectome predicts function (Nature)	Inspiration — we test the same principle at human macro-scale
Brunel	2000	Balanced spiking network theory (J Comput Neurosci)	Our LIF parameter calibration follows this

3. Data Sources

TVB76 — 76 cortical regions

Source: The Virtual Brain project (tvb-data package)
URL: https://raw.githubusercontent.com/the-virtual-brain/tvb-data/master/tvb_data/connectivity/connectivity_76.zip
Contents: 76×76 weights matrix, tract lengths, region labels, 3D coordinates
Bundled at: src/encephagen/connectome/bundled/tvb76_*

TVB96 — 80 cortical + 16 subcortical regions

Source: Same tvb-data package
URL: https://raw.githubusercontent.com/the-virtual-brain/tvb-data/master/tvb_data/connectivity/connectivity_96.zip
Includes: Thalamus (TM-F, TM-T, TM-OP × 2 hemispheres), Basal ganglia (Cd, Pu, Pa, Acc × 2), Amygdala (× 2)
Bundled at: src/encephagen/connectome/bundled/tvb96_*
Primary parcellation used in experiments 2-7

4. Experiments and Results

Phase 1: Wilson-Cowan Oscillators (96 regions)

Experiment 1: Initial Predictions (TVB76)

File: experiments/01_emergent_roles.py
Parameters: Wilson-Cowan, G=0.03, default WC params, identical everywhere
Result: 1/4 predictions supported (P5 regional differentiation, var_cv=1.93). P2 (PFC sustained) marginally significant (p=0.047, ratio=1.04). P1 not testable (no thalamus in TVB76). P4 (frequency differentiation) not significant.

Experiment 2: Comprehensive (TVB96 with subcortical)

File: experiments/02_comprehensive.py
Parameters: Wilson-Cowan (w_ee=16, w_ei=12, w_ie=15, w_ii=3, theta_e=2.0, a_e=1.5), G=0.03
Key results:
- P2 SUPPORTED: PFC tau=4.8ms vs sensory tau=0.6ms (p=0.01, 7.9x ratio)
- P5 SUPPORTED: variance_cv=3.03, tau_cv=1.07, freq_cv=0.50
- Real vs random differentiation: var_cv 1.93 vs 1.81 (p=0.0097)
- Basal ganglia: only region type maintaining oscillations at G=0.03-0.1

Experiment 3: Deep Analysis

File: experiments/03_deep_analysis.py

Key finding — Silencing order:

G=0.015: Thalamus + Motor silence first
G=0.020: Hippocampus + Sensory
G=0.030: Other cortical
G=0.050: Prefrontal
Never:   Basal ganglia

Null model comparison per prediction: PFC tau ratio (real=8.0x, random=7.8x) and BG variance ratio (real=108x, random=120x) are similar — degree distribution explains most.

Experiment 4: Isolate Topology Effects

File: experiments/04_isolate_topology.py
Parameters: G=0.01, 15 null instances per type
Novel finding — Two-level decomposition:
- Level 1 (degree-driven, 24 significant metrics): Oscillation frequencies, variance hierarchy, time constants. Degree-preserving rewiring reproduces these. Erdős-Rényi does not.
- Level 2 (wiring-driven, 11 significant metrics): ALL are functional connectivity patterns. BG-sensory coupling (p=0.0001), hippocampus-motor decoupling (p=0.0001), hippocampus-thalamus coupling (p=0.025), sensory internal coherence (p=0.0005).
- Interpretation: Degree distribution creates the functional hierarchy. Specific wiring creates the communication patterns. Both emerge from identical parameters.

Phase 2: Spiking Neurons (Milestone 1)

E/I Balance Tuning

Problem: LIF neurons either completely silent or fire at 430+ Hz. No intermediate regime.
Root cause: Single synaptic current accumulator mixing E and I with averaged decay.
Solution: Separate i_exc and i_inh tracking. Brunel (2000) calibration: nu_ext = C_ext * nu_thr * ext_rate where nu_thr = V_threshold / (j_eff * C_ext * tau_m).
Auto-scaling: j_eff = j_exc * (16 / C_E) to maintain consistent behavior across population sizes.
Working parameters: LIFParams(j_exc=2.0, g_inh=5.0), ext_rate=3.5, global_coupling=0.05

TVB96 Spiking Hierarchy (19,200 neurons)

Parameters: 200 neurons/region, C_E=16, j_eff=2.0mV, ext_rate=3.5, G=0.05
Result: Mean=11.18 Hz, CV=0.224, biologically realistic

Hierarchy (spiking):

sensory        12.43 Hz  (most active)
other cortical 11.67 Hz
hippocampus    10.26 Hz
thalamus        9.80 Hz
motor           8.74 Hz
basal_ganglia   8.47 Hz
prefrontal      7.46 Hz  (least active)

OPPOSITE of Wilson-Cowan (where BG was most active, sensory silenced)
Spiking hierarchy matches biology better: sensory cortex fires fastest, PFC slowest (consistent with Murray et al. 2014 empirical hierarchy)
Bottom regions: BG-Pa (5.5 Hz), PFCm (6.1 Hz), PHC (6.8 Hz) — exactly the regions known for low spontaneous rates in real brains

Milestone 2: Learning

Experiment 6: Familiarity Effect

File: experiments/06_familiarity_learning.py
Protocol: Present pattern A 30 times with STDP, pattern B 0 times. Test both.
STDP params: a_plus=0.005, a_minus=0.005, tau=20ms, applied every 10 steps
Homeostatic: target_rate=10Hz, tau_homeo=500ms, eta=0.01

Result — HABITUATION DETECTED:

Before learning: A response=2.38 Hz, B response=2.21 Hz (both novel)
After learning:  A response=2.38 Hz, B response=3.01 Hz
Familiarity effect: -0.80 Hz (A weaker than B)

Interpretation: Repetition suppression — the brain responds less to familiar stimuli. This is a well-documented neurological phenomenon (Grill-Spector et al. 2006).
Average weight change: 0.58 mV per region (~29% from baseline j_exc=2.0)
Runtime: 22 minutes for full experiment

Milestone 6: Embodied Learning

Experiment 7: Learn to Stand

File: experiments/07_learn_to_stand.py
Architecture: toy20 connectome, 100 neurons/region, 2000 total neurons
Body: MuJoCo 2-legged walker (4 joints: right/left hip/knee)
Timing: 200 brain steps (20ms) per physics action, 4 physics steps (20ms)
Learning: STDP every 50 brain steps, reward-modulated
Reward: upright bonus + height + forward velocity - tilt - fall penalty
Protocol: 10 baseline (no learning) → 50 learning → 10 test (retention)

Result — BRAIN LEARNED TO STAND:

                      BASELINE    AFTER LEARNING    CHANGE
Avg steps before fall    46.2          50.0          +3.8
Avg reward               46.2          59.9         +13.7
Avg height                0.73          0.81         +0.08
Fall rate                  50%            0%          -50%

Fall rate dropped from 50% to 0%. Reward increased by 30%.
Learning curve: rapid improvement by episode 6-10 (already at 50 steps, 0 falls).

5. Architecture Reference

LIF Neuron Parameters (Working Regime)

LIFParams(
    tau_m=20.0,          # Membrane time constant (ms)
    v_rest=0.0,          # Resting potential (mV, normalized)
    v_threshold=20.0,    # Spike threshold (mV)
    v_reset=0.0,         # Reset after spike (mV)
    t_ref=2.0,           # Refractory period (ms)
    tau_syn_exc=5.0,     # Excitatory synaptic time constant (ms)
    tau_syn_inh=5.0,     # Inhibitory synaptic time constant (ms)
    j_exc=2.0,           # Excitatory PSP amplitude (mV) — for C_E=16
    g_inh=5.0,           # Relative inhibitory strength
)

SpikingBrain Parameters (Working Regime for TVB96)

SpikingBrain(
    connectome=tvb96,
    neurons_per_region=200,     # 19,200 total
    exc_ratio=0.8,              # 80% excitatory, 20% inhibitory
    internal_conn_prob=0.1,     # 10% within-region connectivity
    between_conn_prob=0.02,     # 2% between-region (scaled by connectome weight × G)
    global_coupling=0.05,       # Between-region coupling strength
    ext_rate=3.5,               # External Poisson rate (× Brunel threshold)
)
# Auto-scaled: j_eff = j_exc * (16 / C_E)
# At these params: ~11 Hz mean firing, CV=0.22

STDP Parameters

STDPParams(
    a_plus=0.005,    # LTP amplitude (0.003 for embodied learning)
    a_minus=0.005,   # LTD amplitude
    tau_plus=20.0,   # LTP time window (ms)
    tau_minus=20.0,  # LTD time window (ms)
    w_max=10.0,      # Maximum weight (mV)
    w_min=0.0,       # Minimum weight (mV)
)

Homeostatic Parameters

HomeostaticParams(
    target_rate=10.0,    # Target firing rate (Hz)
    tau_homeo=500.0,     # Slow time constant (ms)
    eta=0.005,           # Scaling rate per adjustment
)

MuJoCo Body

Model: 2-legged walker (torso + 2×thigh + 2×shin + 2×foot)
Joints: right_hip, right_knee, left_hip, left_knee (4 DOF)
Actuators: 4 torque motors, gear=20, range [-1, 1]
Sensors: 4 joint angles, 4 joint velocities, torso pos/quat
Physics dt: 0.005s

6. File Structure

encephagen/
├── src/encephagen/
│   ├── connectome/          # Connectome loading (from cortexlet/conntopo)
│   │   ├── loader.py        # Connectome class
│   │   ├── parcellations.py # Region classification
│   │   └── bundled/         # TVB76 + TVB96 data files
│   ├── dynamics/            # Phase 1: Wilson-Cowan
│   │   ├── wilson_cowan.py  # WC oscillator model
│   │   └── brain_sim.py     # BrainSimulator (WC on connectome)
│   ├── neurons/             # Milestone 1: Spiking
│   │   ├── lif.py           # LIFNeurons (vectorized, separate E/I)
│   │   └── population.py    # RegionPopulation (balanced E/I)
│   ├── network/             # Milestone 1: Brain assembly
│   │   └── spiking_brain.py # SpikingBrain (Brunel-calibrated)
│   ├── learning/            # Milestone 2: Plasticity
│   │   ├── stdp.py          # STDPRule (vectorized, sparse)
│   │   └── homeostatic.py   # HomeostaticPlasticity
│   ├── sensory/             # Milestone 3: Input encoding
│   │   ├── visual.py        # VisualEncoder (images → spikes)
│   │   └── auditory.py      # AuditoryEncoder (audio → spikes)
│   ├── motor/               # Milestone 4: Output decoding
│   │   └── decoder.py       # MotorDecoder (spikes → actions)
│   ├── environment/         # Milestone 4: World
│   │   └── grid_world.py    # GridWorld (2D navigation)
│   ├── body/                # Milestone 6: Physics body
│   │   └── simple_body.py   # SimpleBody (MuJoCo 2-leg walker)
│   ├── loop/                # Milestone 5+6: Closed loop
│   │   ├── closed_loop.py   # ClosedLoopRunner (brain + GridWorld)
│   │   └── embodied_loop.py # EmbodiedLoopRunner (brain + MuJoCo body)
│   ├── analysis/            # Analysis tools
│   │   └── functional_roles.py  # Region classification, predictions
│   └── viz/                 # (placeholder for visualization)
├── experiments/
│   ├── 01_emergent_roles.py       # Phase 1: TVB76 predictions
│   ├── 02_comprehensive.py        # Phase 1: TVB96 + subcortical
│   ├── 03_deep_analysis.py        # Phase 1: Silencing order + null comparison
│   ├── 04_isolate_topology.py     # Phase 1: Degree vs wiring
│   ├── 05_spiking_hierarchy.py    # Milestone 1: Spiking hierarchy test
│   ├── 06_familiarity_learning.py # Milestone 2: STDP habituation
│   └── 07_learn_to_stand.py       # Milestone 6: Embodied learning
├── tests/                   # 81 tests total
│   ├── test_brain_sim.py          # Wilson-Cowan (10 tests)
│   ├── test_functional_roles.py   # Analysis (3 tests)
│   ├── test_spiking.py            # LIF + population + brain (13 tests)
│   ├── test_learning.py           # STDP + homeostatic (14 tests)
│   ├── test_sensory.py            # Visual + auditory (13 tests)
│   ├── test_motor_env.py          # Motor + GridWorld (13 tests)
│   ├── test_body.py               # MuJoCo body (10 tests)
│   ├── test_closed_loop.py        # GridWorld loop (4 tests)
│   └── test_embodied_loop.py      # MuJoCo loop (4 tests)
├── figures/                 # Phase 1 figures
│   ├── fig1_phase_transition.png  # Silencing order
│   ├── fig2_hierarchy_bar.png     # 96 regions ranked
│   └── fig3_fc_comparison.png     # Real vs random FC
├── scripts/
│   └── generate_figures.py        # Figure generation script
├── results/                 # Experiment outputs (gitignored)
├── RDP.md                   # Research Design Proposal v2
├── RDR-2026-04-01.md        # This document
├── README.md                # Project README with figures
└── pyproject.toml           # Package config

7. What's Novel (Publishable Claims)

Claim 1: Two-level decomposition (NOVEL)

Degree distribution determines the functional hierarchy (which regions are active, their time constants). Specific wiring determines functional connectivity patterns (which regions communicate). Separated using degree-preserving rewiring.

Claim 2: Silencing order (NOVEL)

A specific ordering of region-type silencing as coupling increases: thalamus/motor → hippocampus/sensory → other cortex → prefrontal → basal ganglia (never). Not previously reported.

Claim 3: Spiking hierarchy matches biology (EXTENDS prior work)

With identical LIF parameters, the spiking network produces sensory(12.5Hz) > thalamus > motor > BG > PFC(7.5Hz). This matches Murray et al. 2014 empirical hierarchy. Extends Gollo 2015 and Zamora-Lopez 2025 to spiking neurons.

Claim 4: STDP produces habituation (STANDARD but in novel context)

Repetition suppression from STDP on a connectome-structured spiking network. The phenomenon is well-known; the context (human-connectome-topology spiking network) is new.

Claim 5: Embodied learning — INVALIDATED

Experiment 7 claimed the brain learned to stand. Experiment 9 (control) invalidated this:

STDP without reward modulation produces the same result (reward modulation adds nothing)
Motor output torque std drops from 0.26 to 0.03 after STDP (motor death, not active control)
The "0% fall rate" was the body standing stiff, not learned balance

Null Result: Topology doesn't help embodied learning (Experiment 8)

Real connectome vs degree-preserving random wiring for embodied learning: no significant difference (reward 59.6 vs 60.7). The specific wiring pattern does not accelerate motor learning at this scale.

8. Known Limitations and Open Questions

Central unresolved question

Why do Wilson-Cowan and spiking models produce OPPOSITE hierarchies on the same connectome? WC: BG most active, sensory silenced. LIF: sensory most active, BG least active. Until this is explained, "topology drives hierarchy" is not a defensible claim — it should be "the interaction between dynamics model and topology drives hierarchy." (Reviewer 2 feedback)

Limitations

96-region parcellation is coarse. Glasser 360 or Schaefer 200 needed for validation. Results may not hold at finer scales.
Only 15-20 null model instances. Need 500 minimum for robust statistics. Current p-values may not survive.
No empirical FC comparison. Every paper in this tradition compares simulated FC against fMRI. This is conspicuously absent.
Chaudhuri et al. (2015) is a challenge, not a predecessor. They showed that removing heterogeneous parameters eliminates the hierarchy — directly contradicting our "topology alone" claim. Must be framed as such.
Degree-preserving rewiring preserves more than degree. Also preserves assortativity, rich-club structure. Need rich-club-specific null models to truly isolate degree.
Spiking E/I balance breaks at C_E=80. j_eff auto-scaling is empirical, not principled. 200 neurons/region (C_E=16) is a different connectivity regime from biology (~C_E=1000).
No conduction delays. At 6 m/s, a 120mm tract introduces 20ms delay — same timescale as STDP window.
STDP learning rate too high. 29% weight change from 30 presentations is excessive. Likely an artifact.
Familiarity effect has no statistical test. 0.80 Hz effect, N=1, no permutation test, no error bars.
Embodied learning doesn't work. STDP produces motor death (constant torque output), not active motor control. Basic STDP is insufficient for motor learning — need e-prop or three-factor rules.
Performance too slow for interactive use. 640x slower than real-time. Need Norse/NEST for Milestone 5+.
"Miniature human brain" framing is not defensible. No cell-type diversity, no laminar structure, no neuromodulation, no dendritic computation. "Connectome-topology-driven emergence in spiking networks" is accurate.

9. Expert Reviewer Feedback (2 rounds, 4 reviewers)

Two rounds of expert panel review were conducted. Full feedback is documented separately. Key demands:

Must do (blocks credibility)

Explain WC/spiking hierarchy inversion — this is the central scientific question
Run two-level decomposition on spiking network (not just WC)
Compare simulated FC against empirical HCP fMRI FC (Pearson's r, upper triangle)
Increase null models to ≥500 instances
Reframe Chaudhuri 2015 as "challenge" not "extension"
Retire "miniature human brain" framing
Remove/qualify OpenWorm/FlyWire analogy

Should do (strengthens science)

Add rich-club-preserving null model
Statistical tests for familiarity: 20 runs with different seeds, permutation test
Reduce STDP a_plus by 10x, verify habituation survives
Compute autocorrelation timescales from spike trains (not just firing rates)

For embodied learning (if pursued)

Random connectome control (done — Exp 8, no difference)
STDP without reward control (done — Exp 9, reward adds nothing)
Frozen weights control (not done)
10x longer training (not done)
Scale to TVB96 (not done, blocked by C_E=80 calibration)

10. Next Steps for a New Researcher

Priority 1: The central question (determines if this is publishable)

Explain the WC/spiking hierarchy inversion. Is it a WC sigmoid artifact? A Brunel calibration artifact? An operating-point-dependent interaction with topology? This requires a dedicated analysis, not a footnote.
Run two-level decomposition on spiking network. Does degree-preserving rewiring reproduce the LIF hierarchy? If yes, the finding extends to a more biologically plausible substrate.

Priority 2: Statistical rigor

Increase null models to 500 instances. Current 15-20 is a pilot.
Compare simulated FC vs empirical HCP fMRI. Download HCP resting-state FC, compute Pearson's r. This is the standard validation in the field.
Run familiarity experiment 20 times with different seeds. Compute mean ± SD, permutation test.

Priority 3: Learning

Replace basic STDP with e-prop (Bellec et al. 2020) or three-factor rules (Frémaux & Gerstner 2016). Basic STDP is insufficient for motor control.
Migrate to Norse (PyTorch spiking) or NEST for GPU acceleration. Current NumPy is too slow.

Priority 4: Scaling

Fix E/I balance at C_E=80 for 1000 neurons/region on TVB96.
Test with Schaefer 200 parcellation to verify parcellation independence.
Add conduction delays from TVB tract lengths.

11. How to Reproduce Everything

# Clone and install
git clone https://github.com/toroleapinc/encephagen.git
cd encephagen
pip install -e ".[dev]"
conda install -c conda-forge mujoco  # For embodied experiments

# Run all tests (81 tests, ~10 minutes)
python -m pytest tests/ -v

# Phase 1 experiments (Wilson-Cowan)
python experiments/01_emergent_roles.py    # ~2 min
python experiments/02_comprehensive.py     # ~5 min
python experiments/03_deep_analysis.py     # ~15 min
python experiments/04_isolate_topology.py  # ~10 min

# Spiking hierarchy
python experiments/05_spiking_hierarchy.py # ~30 min

# Learning
python experiments/06_familiarity_learning.py  # ~22 min

# Embodied (includes negative results and controls)
python experiments/07_learn_to_stand.py         # ~15 min (initial claim)
python experiments/08_topology_vs_random_embodied.py  # ~50 min (null result)
python experiments/09_learning_control.py       # ~30 min (invalidates Exp 7)

# Generate Phase 1 figures
python scripts/generate_figures.py

Hardware Used

GPU: NVIDIA RTX 5070 (12 GB VRAM) — not used for computation (all NumPy/CPU)
CPU: Intel Core Ultra 7 265 (20 cores)
RAM: 16 GB
OS: WSL2 on Windows, Ubuntu

10. How to Reproduce Everything

# Clone and install
git clone https://github.com/toroleapinc/encephagen.git
cd encephagen
pip install -e ".[dev]"
conda install -c conda-forge mujoco  # For Milestone 6

# Run all tests (81 tests, ~10 minutes)
python -m pytest tests/ -v

# Phase 1 experiments (Wilson-Cowan)
python experiments/01_emergent_roles.py    # ~2 min
python experiments/02_comprehensive.py     # ~5 min
python experiments/03_deep_analysis.py     # ~15 min
python experiments/04_isolate_topology.py  # ~10 min

# Milestone 1: Spiking hierarchy
python experiments/05_spiking_hierarchy.py # ~30 min

# Milestone 2: Familiarity learning
python experiments/06_familiarity_learning.py  # ~22 min

# Milestone 6: Embodied learning
python experiments/07_learn_to_stand.py    # ~15 min

# Generate Phase 1 figures
python scripts/generate_figures.py

Hardware Used

GPU: NVIDIA RTX 5070 (12 GB VRAM) — not used for computation (all NumPy/CPU)
CPU: Intel Core Ultra 7 265 (20 cores)
RAM: 16 GB
OS: WSL2 on Windows, Ubuntu

11. Sibling Projects

cortexlet (https://github.com/toroleapinc/cortexlet) — Brain-topology-structured trainable PyTorch neural network. 98% MNIST accuracy, 45 tests. Not novel as ML contribution.
conntopo (https://github.com/toroleapinc/conntopo) — Toolkit for connectome dynamics vs null models. Reproduces established findings (Honey 2007+). 27 tests. Value as open-source tool, not as discovery.

FilesExpand file tree

RDR-2026-04-01.md

Latest commit

History

RDR-2026-04-01.md

File metadata and controls

Encephagen: Research Development Report

1. Project Overview

2. Prior Work (Must Cite in Any Paper)

3. Data Sources

TVB76 — 76 cortical regions

TVB96 — 80 cortical + 16 subcortical regions

4. Experiments and Results

Phase 1: Wilson-Cowan Oscillators (96 regions)

Experiment 1: Initial Predictions (TVB76)

Experiment 2: Comprehensive (TVB96 with subcortical)

Experiment 3: Deep Analysis

Experiment 4: Isolate Topology Effects

Phase 2: Spiking Neurons (Milestone 1)

E/I Balance Tuning

TVB96 Spiking Hierarchy (19,200 neurons)

Milestone 2: Learning

Experiment 6: Familiarity Effect

Milestone 6: Embodied Learning

Experiment 7: Learn to Stand

5. Architecture Reference

LIF Neuron Parameters (Working Regime)

SpikingBrain Parameters (Working Regime for TVB96)

STDP Parameters

Homeostatic Parameters

MuJoCo Body

6. File Structure

7. What's Novel (Publishable Claims)

Claim 1: Two-level decomposition (NOVEL)

Claim 2: Silencing order (NOVEL)

Claim 3: Spiking hierarchy matches biology (EXTENDS prior work)

Claim 4: STDP produces habituation (STANDARD but in novel context)

Claim 5: Embodied learning — INVALIDATED

Null Result: Topology doesn't help embodied learning (Experiment 8)

8. Known Limitations and Open Questions

Central unresolved question

Limitations

9. Expert Reviewer Feedback (2 rounds, 4 reviewers)

Must do (blocks credibility)

Should do (strengthens science)

For embodied learning (if pursued)

10. Next Steps for a New Researcher

Priority 1: The central question (determines if this is publishable)

Priority 2: Statistical rigor

Priority 3: Learning

Priority 4: Scaling

11. How to Reproduce Everything

Hardware Used

10. How to Reproduce Everything

Hardware Used

11. Sibling Projects