Capability Graph Neuroscience Validation

Experimental validation of five neuroscience-inspired learning mechanisms (STDP, homeostatic scaling, BCM metaplasticity, engram drift, neural recruitment) implemented in a Hebbian capability graph. 18 controlled experiments, 4 falsified hypotheses, 1 production graph analysis.

Paper draft: PAPER_DRAFT.md

---

What we found

A production capability graph with 73 nodes and 357K activations reached 99.8% edge weight saturation -- every edge at maximum, no discrimination between frequent and rare connections. The graph had remembered everything and could distinguish nothing.

We implemented six forgetting mechanisms drawn from neuroscience (temporal decay, homeostatic scaling, Hebbian unlearning, consolidation by similarity, stale edge pruning, contextual inhibition) and validated each through controlled experiments.

Strongest result: Removing STDP drops directional sequence discrimination from 100% to 48% (chance). Cohen's d = 3.163, p < 0.000001. Hebbian learning alone detects co-occurrence but cannot distinguish A->B from B->A.

Validated mechanisms:

• STDP provides 100% of directional awareness (ablation: 100% -> 48%)
• Homeostatic scaling prevents saturation (ablation: 14.45% ceiling without decay)
• BCM strengthens encoding by 58% (ablation: -58% encoding strength)
• Drift tracking confirmed as pure observer (zero learning impact, as designed)
• Neural recruitment improves bridge discovery by 121% (Cohen's d = 19.8)

Falsified hypotheses:

• Classical STDP timescales do not apply (tau = 999.5s vs. biological 20ms)
• Weight distribution is unimodal, not bimodal (no w_min clamp in source)
• Timescale ordering inverted from prediction
• Network topology is hub-and-spoke, not small-world (C_ratio = 1.19)

---

Repository contents

capability-graph-neuroscience/
  PAPER_DRAFT.md              Paper draft with full methodology and references
  experiments/                18 experiments, each with:
    experiment_NN_*.py          Python script (reproducible, seed=42)
    experiment_NN_data.json     Raw measurements
    experiment_NN_plot.png      6-panel visualization
    experiment_NN_report.md     Full analysis report
  docs/                       Additional documentation
  src/                        Source modules under validation
  analysis/                   Analysis scripts
  results/                    Consolidated results

---

Experiments

STDP Learning

#	Test	Result	Key metric
1	STDP kernel shape	PASS	Exact 2:1 potentiation bias (R^2 = 1.0)
2	STDP temporal precision	FAIL	tau = 999.5s, not classical 20s

Homeostatic Scaling

#	Test	Result	Key metric
3	Weight distribution	FAIL	Unimodal, not bimodal
4	Convergence rate	PASS	0.9 decay rate (R^2 = 1.0)

BCM Metaplasticity

#	Test	Result	Key metric
5	Per-edge theta dynamics	PASS	67% per-edge variance (live data)
6	Consolidation effect	PASS	15.5x learning rate difference

Engram Drift

#	Test	Result	Key metric
7	Pattern persistence	PASS	All constellations 100-160x below threshold
8	Drift velocity	PASS	Formula verified exactly (R^2 = 1.0)

System Integration

#	Test	Result	Key metric
9	Perturbation recovery	PASS	100% recovery in mean 79 events
10	Multi-timescale	FAIL	Timescale ordering inverted
11	Small-world topology	FAIL	Core-periphery (C_ratio = 1.19)

Ablation Studies

#	Test	Result	Key metric
12	Full system baseline	PASS	Convergence at event 1050
13	Remove STDP	PASS	Order discrimination: 100% -> 48%
14	Remove homeostatic	PASS	14.45% ceiling saturation
15	Remove BCM	PASS	-58% encoding strength
16	Remove drift	PASS	Zero learning impact (pure observer)

Production + Recruitment

#	Test	Result	Key metric
17	Production graph analysis	PASS	8 critical discoveries including edge saturation
18	Neural recruitment ablation	PASS	121% bridge discovery improvement (d = 19.8)

Overall: 13/18 PASS, 4/18 FAIL, 1 production analysis

---

Reproducing

Requires Python 3.12, NumPy, SciPy, Matplotlib.

cd experiments/
python experiment_01_stdp_kernel.py

Each script writes its own _data.json, _plot.png, and _report.md. Seeds are fixed at 42.

---

Stack

• Python 3.12
• NumPy / SciPy / Matplotlib
• Soul Matrix: Rust (512x512 Hebbian conductance matrix, ~112us inference)
• Capability Graph: Python (Hebbian + STDP + homeostatic scaling + BCM + recruitment)

---

Authors

CIPS Corp LLC -- February 2026

Contact: glass@cipscorps.io

---

License

MIT