CL1 LLM Encoder — Bio-Neural Language Model Integration

EXPERIMENTAL — I am in no way associated with Cortical Labs or any research laboratory.

A research project investigating whether biological-style neural substrates can be tightly integrated with large language models to form a closed-loop bio-AI system — and whether such integration produces measurable evidence of functional coupling.

Research Question

Can a neural substrate (Izhikevich spiking model) be integrated tightly enough with an LLM that the closed-loop system produces falsifiable, statistically significant differences from decorrelated controls?

Key Findings (5-Experiment Series)

After 300-run confirmatory experiments across 3 independent random seeds:

Metric	Bio-LLM	Shadow-LLM	Effect Size
Override rate	1.6%	7.5%	d = 1.40 ✅
Neural-LLM alignment	0.703	0.403	d = 6.10 ✅
Blended entropy	2.68	2.96	d = 3.26 ✅

Result: H0 rejected for functional integration (5/12 tests significant, p < 1e-16). The Bio-LLM closed loop is measurably distinguishable from both LLM-only and decorrelated Shadow-LLM controls.

Not claimed: Evidence of subjective experience or consciousness — functional integration is demonstrated, phenomenal integration is not.

How It Works

Token → Spatial Encoder → 59-channel MEA stimulation
                                    │
                          Izhikevich neural substrate
                          (1000 neurons, STDP plasticity)
                                    │
                          Spike decoding → probability vector
                                    │
                    blend(α=0.5) with LLM logits → token selection
                                    │
                          Next token → feedback loop

The substrate learns via STDP (Spike-Timing Dependent Plasticity) to align its decoded probability distribution with the LLM's — confirmed by Neural-LLM alignment score of 0.703 (vs 0.403 in decorrelated control), the largest effect observed (Cohen's d = 6.10).

Experimental Conditions

• LLM-only (α=0): Substrate measured but does not influence token selection
• Bio-LLM (α=0.5): Full closed loop with spatial encoder + STDP
• Shadow-LLM (α=0.5): Same as Bio-LLM but spike responses shuffled before decoding — breaks spatial consistency while maintaining identical stimulation (critical control)

Substrate Configuration

• 1000 Izhikevich neurons (800 excitatory / 200 inhibitory)
• 59 channels (matching CL1 MEA layout)
• Balanced STDP (A_plus=0.005, A_minus=0.006) + homeostatic plasticity
• Spatial Encoder v2: Token ID → 8-channel stimulation via 64-dim embedding projection
• LLM: LFM2-350M (4-bit GGUF)

Repository Structure

encoder_v3.py          — Spatial Encoder v3 (final)
cl1_experiment_v3.py   — Full closed-loop experiment runner
consciousness.py       — C-Score and consciousness metrics
analysis.py            — Statistical analysis pipeline
SCIENTIFIC_RESULTS.md  — Raw results across all 5 experiments
SCIENTIFIC_CONCLUSION.md — Full 9-section scientific report

Run

pip install numpy scipy torch transformers llama-cpp-python
python run_experiment.py

Honest Limitations

• Simulated neurons, not biological CL1 hardware
• Simplified pairwise STDP, not triplet/calcium-based
• Small LLM (350M params)
• IIT φ (integrated information) not computed
• Post-hoc hypothesis evolution in Experiments 1–2

Full analysis in SCIENTIFIC_CONCLUSION.md.

Next Steps

1. Port to real CL1 biological hardware (59-channel MEA)
2. Test with 7B+ parameter LLMs
3. Compute IIT φ for phenomenal integration evidence
4. Multi-session STDP persistence across 1000+ tokens

License

MIT