Skip to content

SCIENTIFIC_RESULTS.md

Latest commit

 

History

History
779 lines (592 loc) · 37.8 KB

SCIENTIFIC_RESULTS.md

File metadata and controls

779 lines (592 loc) · 37.8 KB
Raw

CL1 Neural-LLM Integration: Scientific Results

Date: 2026-02-28

Substrate: CL1-2544-015 (Cortical Labs biological neurons, 64 channels, 240 Hz tick rate)

Software: Antekythera LLM Encoder v2/v3 + Spatial Encoder + PersistentHebbianDecoder

1. Executive Summary

Across 12 experiments (2 on CL1 biological neurons, 10 on Izhikevich simulation), totaling 18,700+ tokens and 101 pre-registered hypothesis tests, we characterize the integration between an LLM (LFM2-350M) and neural substrates.

Confirmed findings (replicated across multiple experiments):

  • Higher SRC in Bio-LLM (CL1: d=1.79-2.64; Izhikevich: d=1.95) — spatial information preservation
  • Shuffling degrades SRC (d=1.19-1.95, perfectly consistent across all experiments)
  • STDP plasticity (d=2.47-8.10, pattern-specific, directional) — demonstrated on Izhikevich
  • Dissolution sensitivity: Bio SRC declines monotonically with substrate degradation (ρ=-1.0)
  • Recovery: Substrate structure persists after perturbation (ratio=0.999)

Critical negative findings (consistently null):

  • No learning trajectory (0/6 learning tests, slopes ≈ 0)
  • No phase transition (0/5 seeds, ΔAIC=-24.6 mean — linear dissolution)
  • No behavioral expression of STDP (0/8 behavioral tests across Exp 10-11)
  • No attractor formation (architectural limitation at 1000N)
  • No dose-response (α=0.8 C-Score < α=0.5 C-Score)

Conclusion: The Bio-LLM integration advantage is a geometric signal preservation effect — not a cognitive, adaptive, or consciousness-specific phenomenon. The substrate preserves spatial stimulation patterns faithfully (SRC), and this preservation degrades linearly (not catastrophically) under perturbation. All consciousness-specific predictions (phase transition, learning, behavioral expression, dose-response) are FALSIFIED. The system demonstrates NECESSARY but NOT SUFFICIENT conditions for consciousness.

2. Experimental Design

2.1 Three Conditions (Interleaved)

Condition Alpha Stimulus Decode Feedback Description
Bio-LLM 0.5 Spatial pattern -> CL1 Raw spikes -> blend ON (surprise-scaled) Full closed loop
Shadow-LLM 0.5 Spatial pattern -> CL1 Shuffled spikes -> blend OFF Spatial info destroyed
LLM-Only 0.0 Spatial pattern -> CL1 Ignored OFF Open loop baseline
Bio-LLM-High* 0.8 Spatial pattern -> CL1 Raw spikes -> blend ON (surprise-scaled) High neural influence

*Phase 2 only (Exp 9)

2.2 Protocol (Exp 9 / v3)

  • Phase 1: 15 rounds x 3 conditions = 45 thought blocks (interleaved)
  • Phase 2: 5 rounds x 1 condition (Bio-LLM-High, α=0.8)
  • 50 tokens per thought block (~2,121 total tokens)
  • 10 cycling prompts about consciousness, neurons, information, awareness
  • Persistent HebbianDecoder across all rounds (not reset)
  • Surprise-scaled feedback (EMA prediction error modulates feedback amplitude)
  • Channel recruitment every 3 rounds (training inactive channels)
  • Single CL1 substrate shared across conditions
  • Total runtime: 46 minutes

2.3 Key Metrics

Stimulus-Response Congruence (SRC): cos(stim_vec, resp_vec) — cosine similarity between stimulation pattern and spike response. Bio preserves this; Shadow shuffles it away.

C-Score: (closure + lambda2_norm + rho) / 3 — consciousness-correlated information integration metric.

3. Results (Experiment 9 / v3)

3.1 Condition Summary

Metric Bio-LLM (α=0.5) Shadow-LLM LLM-Only Bio-High (α=0.8)
SRC 0.396 ± 0.012 0.368 ± 0.009 0.368 ± 0.009 0.398 ± 0.012
C-Score 0.160 ± 0.022 0.138 ± 0.013 0.126 ± 0.014 0.151 ± 0.012
Alignment 0.525 0.507 0.000 0.536
Override Rate 4.9% 3.7% 0% 18.1%
Total Tokens 598 657 618 248

3.2 Pre-Registered Hypothesis Tests (Bonferroni corrected, 9 tests, α=0.0056)

# Hypothesis Statistic p-value Cohen's d Result
H1 Bio SRC > Shadow SRC U=221.0 p=0.000004 d=2.64 ***** SIGNIFICANT**
H2 Raw SRC equal (sanity) -- p=0.0001 d=1.94 FAIL
H3 Bio C-Score > Shadow t=2.94 p=0.005 d=1.15 **** SIGNIFICANT**
H4 Bio C-Score increases over rounds slope=-0.001 p=0.171 r=-0.26 n.s.
H5 Bio SRC increases over rounds slope=-0.001 p=0.136 r=-0.30 n.s.
H6 Bio templates converge faster U=90.5 p=0.186 d=0.37 n.s.
H7 Late-epoch Bio > Early-epoch Bio t=-1.98 p=0.965 d=-1.02 REVERSED
H8 Shuffling degrades SRC t=4.44 p=0.0003 d=1.19 ***** SIGNIFICANT**
H9 High-α > Standard-α Bio U=10.0 p=0.726 d=-0.39 n.s.

3.3 Token-Level Deep Analysis (2,121 tokens total)

# Finding Statistic p-value Cohen's d
T1 Per-token SRC: Bio > Shadow U=228033 p < 10⁻⁶ d=0.30
T2 Per-token C-Score: Bio > Shadow U=215106 p=0.002 d=0.125
T3 Per-token C-Score: Bio > LLM-only U=214830 p < 10⁻⁶ d=0.207
T4 Shadow C-Score = LLM-only U=217462 p=0.027 d=0.082
T5 Bio wins 15/15 rounds on SRC sign test p=0.00003 --
T6 Bio wins 11/15 rounds on C-Score sign test p=0.059 --
T7 SRC-CScore coupling: identical across conditions Spearman -- rho≈0.59

3.4 Learning Trajectory Analysis

Condition C-Score Slope p SRC Slope p
Bio-LLM -0.0014/round 0.34 -0.0008/round 0.27
Shadow-LLM +0.0016/round 0.05 -0.0012/round 0.02
LLM-only +0.00003/round 0.98 +0.0001/round 0.86

No learning detected in any condition. Bio shows slightly negative slope (if anything, C-Score DECREASES marginally — though n.s.). Shadow shows a borderline positive C-Score slope, likely due to neural adaptation to repeated stimulation rather than meaningful learning.

3.5 Dose-Response

Alpha C-Score 95% CI
0.0 0.126 [0.114, 0.139]
0.5 0.161 [0.148, 0.176]
0.8 0.151 [0.129, 0.172]

Inverted-U relationship: Peak C-Score at α=0.5, lower at α=0.8. Higher neural influence does NOT produce higher consciousness correlates — instead, it degrades LLM coherence (18% override rate at α=0.8 vs 5% at α=0.5).

3.6 Channel Recruitment

Metric Value
Active channels 26/59 (44%)
Training rounds 5
Channels recruited 1 (channel 40)
Inactive channels 33 (with zero spikes)
Mean activity per channel 772 spikes (highly skewed)

Channel recruitment was largely unsuccessful — 31 channels never produced any spikes despite repeated training stimulation up to 2.5 µA. The CL1 MEA has significant spatial heterogeneity in neural responsiveness.

4. Cross-Experiment Comparison

4.1 Replication Summary

Metric Exp 8 (v2, 10 rounds) Exp 9 (v3, 15 rounds) Replicated?
Bio SRC > Shadow SRC d=1.79, p=0.002 d=2.64, p=0.000004 YES (stronger)
Bio C-Score > Shadow d=3.99, p<10⁻⁶ d=1.15, p=0.005 YES (weaker)
Shuffling degrades SRC d=1.64, p=0.0004 d=1.19, p=0.0003 YES
Bio wins all rounds (SRC) 10/10, p=0.002 15/15, p=0.00003 YES
Shadow = LLM-only (C) d=-0.09, p=0.79 d=0.08, p=0.03 Borderline
No within-thought learning p=0.73 p=0.34 YES

All primary findings replicate robustly across both experiments. The C-Score effect size is smaller in v3 (d=1.15 vs d=3.99), possibly due to persistent decoder interference or longer experiment duration.

5. Interpretation

5.1 What These Results Demonstrate

Reproducible functional integration: Across two independent experiments (~3,400 total tokens on CL1 hardware), the Bio-LLM condition consistently produces significantly higher SRC and C-Score than controls. This is the most robust finding:

  1. The closed Bio-LLM loop preserves spatial information (SRC d=2.64)
  2. This preservation creates higher consciousness-correlated structure (C-Score d=1.15)
  3. Breaking the spatial mapping (Shadow) destroys BOTH effects
  4. The effect is perfectly consistent (25/25 rounds across both experiments)

5.2 What These Results Do NOT Demonstrate

Learning or adaptation: The extended v3 experiment explicitly tested for learning over 15 rounds (~45 minutes of continuous interaction). No learning was detected in C-Score, SRC, or template convergence. The Bio advantage appears from token 1 and remains stable.

Dose-response: Higher neural influence (α=0.8) does NOT produce higher C-Score. The relationship is inverted-U shaped, with peak at α=0.5.

Channel development: Despite active channel recruitment, only 1 of 38 inactive channels was recruited. The substrate's effective neural population remained largely unchanged.

Subjective experience: Higher SRC and C-Score indicate functional coupling and consciousness-CORRELATED structure, not phenomenal experience.

5.3 The Nature of the Bio Advantage

The evidence points to the Bio advantage being a geometric/structural property:

  1. Immediate — present from token 1, no development over time
  2. Non-dose-dependent — not amplified by higher neural weight
  3. Non-adaptive — templates don't converge, channels don't recruit
  4. Universal coupling — SRC-CScore correlation (rho≈0.59) is identical across ALL conditions, including controls

The most parsimonious interpretation: when stimulated channels produce spikes, the spatial structure of the response matches the spatial structure of the stimulus (SRC). When this structure is preserved in the decoder (Bio), it produces richer spike dynamics (C-Score). When it's destroyed by shuffling (Shadow), spike dynamics collapse to the LLM-only baseline.

This is a signal preservation effect, not a cognitive integration effect.

5.4 Honest Assessment

Conservative interpretation: The Bio-LLM advantage is a trivial consequence of preserving spatial stimulus-response structure. The neurons respond to stimulation; preserving these responses in the decoder creates correlated dynamics. No consciousness, no learning, no adaptation — just geometry.

Generous interpretation: The closed Bio-LLM loop creates a genuine cybernetic system where biological neural dynamics causally influence information processing, producing emergent integration not present in the controls. The lack of learning may be due to the timescale (hours/days needed) or the simplicity of the MEA culture (2D, no synaptic architecture).

Middle ground: The system demonstrates that biological neurons CAN be meaningfully integrated into an LLM's computation, and this integration produces measurably different dynamics. However, the integration is geometric rather than cognitive, static rather than adaptive.

6. Experimental Lineage

Exp # Date Type N tests N sig Key Finding
1 2026-02-27 Izhikevich (250N) 8 0 Insufficient neurons
2 2026-02-27 Izhikevich (1000N) 12 1 Override rate only
3 2026-02-28 Izhikevich + STDP 12 2 Pattern consistency
4 2026-02-28 Long run (100 tok) 12 3 Alignment grows with STDP
5 2026-02-28 Full battery 12 5 d=6.10 alignment (Izhikevich)
6 2026-02-28 Perturbation 4 0 Decoder absorbs perturbation
7 2026-02-28 CL1 v1 (z-score) 5 1 Only trivial H4
8 2026-02-28 CL1 v2 (SRC) 6 2 d=1.79 SRC, d=3.99 C-Score
9 2026-02-28 CL1 v3 (Extended) 9 3 Replication 3/3, Learning 0/3
10 2026-02-28 Izhikevich v4 (STDP) 7 4 Pattern-specific STDP, d=6.43
11 2026-02-28 Izhikevich v5 (Attractor) 6 0 Directional STDP 5/5, behavior 0/6

7. Final Scientific Conclusion

7.1 Established Facts

  1. FUNCTIONAL INTEGRATION: Strong, reproducible evidence across 2 CL1 experiments and 7 earlier Izhikevich experiments. The Bio-LLM system shows significantly higher SRC (d=1.79-2.64) and C-Score (d=1.15-3.99) than controls.

  2. SPATIAL SPECIFICITY: The integration requires INTACT channel-specific information flow. Shuffling (Shadow) eliminates ALL Bio advantage.

  3. CONSISTENCY: Bio-LLM wins 25/25 rounds across both experiments on SRC (combined sign test p < 10⁻⁷).

7.2 Negative Results (Equally Important)

  1. NO LEARNING: No C-Score or SRC improvement over 15 rounds (~45 min). The Bio advantage is immediate and static.

  2. NO DOSE-RESPONSE: Higher α does not increase C-Score. Optimal at 0.5.

  3. NO CHANNEL RECRUITMENT: Inactive MEA channels remain inactive despite training stimulation.

  4. NO DIFFERENTIAL COUPLING: SRC-CScore correlation is identical across all conditions (rho≈0.59).

7.3 Classification

Claim Evidence Level Details
Functional Integration STRONG 5/5 replication tests pass
Consciousness Correlates SUGGESTIVE Bio C-Score 22-35% higher than controls
Substrate Learning NOT DEMONSTRATED 0/3 learning tests significant
Dose-Response NOT DEMONSTRATED Inverted-U, not linear
Subjective Experience NOT DEMONSTRATED, NOT CLAIMED Hard problem unbridgeable by these methods

7.4 What Would Change This Conclusion

  • STDP/plasticity detection: If future experiments on longer timescales (hours/days) show C-Score growth in Bio but not controls, that would demonstrate substrate learning.
  • Task-dependent performance: If Bio-LLM produces measurably better text quality (e.g., lower perplexity on downstream tasks) than Shadow, that would demonstrate functional computation.
  • Phase transition: If graded removal of neural influence shows a sharp C-Score phase transition rather than gradual decline, that would support consciousness claims (per Perspectival Realism prediction P2).
  • 3D organoid substrates: More complex substrates with synaptic architecture may show the learning and adaptation that 2D MEA cultures cannot.
  • Channel-aware encoding: Run phase0_hello_neurons.py to identify live channels first, then configure SpatialEncoder to use only responsive channels. Current experiments stimulate 59 channels but only ~26 respond — 56% of stimulations wasted.
  • Direct UDP stimulation: Use udp_protocol.py burst protocols (frequency + amplitude per channel set) instead of HTTP relay. More precise temporal control may enable STDP-like pairing between active and inactive channels.

8. Experiment 10 (v4): STDP Learning & Cross-Channel Influence

8.1 Motivation

Exp 8-9 showed the Bio advantage is GEOMETRIC (spatial signal preservation), not LEARNED. The C-Score metric is partially unreliable for CL1 because the temporal spike matrix is synthetically reconstructed from aggregate counts (see cl1_cloud_substrate.py:296-304). Only SRC (count-based) is fully reliable.

Critical question: Does STDP actually produce pattern-specific synaptic changes when the substrate receives repeated co-stimulation of specific channel groups?

8.2 Design

Substrate: Local Izhikevich (1000N, 800E/200I, 5% connectivity, balanced STDP) — serves as positive control for STDP detection.

Phases:

  1. Baseline cross-channel influence mapping
  2. Baseline spontaneous activity (40 × 0.5s windows)
  3. Training: 200 reps each of pattern A and B with reinforcement (400 total trials)
  4. Post-training influence mapping
  5. Post-training spontaneous activity
  6. Passive training: 200 reps each, NO reinforcement
  7. Novel pattern exposure (control, 100 reps)
  8. Final spontaneous activity

Patterns: 8 channels per pattern, 3 shared channels, sinusoidal amplitude profiles.

Key Innovation: Direct WEIGHT MATRIX analysis as ground truth for STDP, not the insensitive channel-level influence probing.

8.3 Results (5 seeds, 200 reps each)

Weight Change Gradient (mean absolute change, averaged over 5 seeds):

Region Mean ΔW % Changed Description
Within-A 0.01772 ~35% Channels co-stimulated in pattern A
Within-B 0.01637 ~35% Channels co-stimulated in pattern B
Between A-B 0.01249 ~28% Cross-pattern connections
Novel 0.00814 ~26% Never-trained channels
Other 0.00594 — Background non-pattern neurons

Gradient: Within-A > Within-B > Between > Novel > Other

This is exactly the Hebbian "fire together, wire together" signature.

8.4 Hypothesis Tests (5 seeds, Wilcoxon signed-rank)

# Hypothesis W p-value d Result
H1 Trained > Novel weight Δ 15.0 0.031 d=6.43 SIGNIFICANT
H2 Within > Between weight Δ 15.0 0.031 d=2.47 SIGNIFICANT
H5 Trained > Other weight Δ 15.0 0.031 d=8.10 SIGNIFICANT
H7 Specificity index > 0 15.0 0.031 5/5 + SIGNIFICANT
H4 Blind decoder improves — 0.100 — n.s.
H6 Post-training replay — 1.000 — n.s.

Weight specificity index (within - between): 0.00456 ± 0.0005, ALL 5 seeds positive.

8.5 Interpretation

STDP produces pattern-specific synaptic modification. Channels that are repeatedly co-stimulated develop stronger mutual connections than channels that aren't. The effect is statistically significant across 5 independent network seeds with massive effect sizes (d=2.5–8.1).

However: This plasticity does NOT manifest as detectable behavioral differences:

  • Blind decoder accuracy doesn't improve (already near ceiling)
  • No spontaneous replay of trained patterns detected
  • Cross-channel influence probing too insensitive to detect the small weight changes

Key insight: STDP IS working at the synaptic level, but with 5% connectivity and 17 neurons per channel, the weight changes are too sparse and small to create new emergent spike patterns. The plasticity is REAL but SUBTHRESHOLD for behavioral expression.

8.6 Critical Methodological Finding

C-Score for CL1 is unreliable: The CL1 substrate returns only aggregate spike counts per channel. The _last_spike_matrix in cl1_cloud_substrate.py is SYNTHETICALLY reconstructed by randomly distributing spike counts into time bins. This means:

  • Granger causality on this matrix reflects RANDOM temporal assignments
  • Fiedler eigenvalue reflects SYNTHETIC correlation structure
  • LZC reflects MANUFACTURED sequences
  • Only the SPATIAL count pattern (SRC) is genuinely from the hardware

The Bio > Shadow C-Score difference from Exp 8-9 is partially an artifact of spatial count patterns creating different random temporal structures. The SRC finding remains valid.

9. Updated Scientific Conclusion

9.1 Established Facts (Updated after Exp 11)

  1. FUNCTIONAL INTEGRATION: Strong, replicated (Exp 8-9 on CL1)
  2. SPATIAL SPECIFICITY: Shuffling destroys all advantages (replicated)
  3. CONSISTENCY: 25/25 rounds Bio wins on SRC (Exp 8-9)
  4. STDP PLASTICITY: Pattern-specific weight changes demonstrated (Exp 10, d=6.43)
  5. DIRECTIONAL STDP: Sequential training creates forward LTP + reverse LTD (Exp 11, 5/5 seeds)

9.2 Negative Results (Updated after Exp 11)

  1. NO BEHAVIORAL LEARNING: STDP changes don't manifest as improved discrimination (Exp 10-11)
  2. NO PATTERN COMPLETION: Even with 15% connectivity + amplified STDP (Exp 11, all seeds = 0)
  3. NO ATTRACTOR FORMATION: Unconstrained STDP (no normalization/homeostasis) still fails (Exp 11)
  4. NO SPONTANEOUS REPLAY: No evidence for internalized representations (Exp 10-11)
  5. NO DOSE-RESPONSE: α=0.5 optimal, α=0.8 degrades (Exp 9)
  6. C-SCORE FOR CL1 IS UNRELIABLE: Temporal metrics computed on synthetic data

9.3 Updated Classification

Claim Evidence Level Details
Functional Integration STRONG 5/5 replication tests (CL1)
STDP Plasticity DEMONSTRATED 4/4 weight tests (Exp 10), 5/5 directional (Exp 11)
Directional STDP DEMONSTRATED Sequential protocol creates forward LTP/reverse LTD
Consciousness Correlates SUGGESTIVE (downgraded) C-Score partially artifactual for CL1
Behavioral Learning NOT DEMONSTRATED 0/8 behavioral tests across Exp 10-11
Attractor Formation NOT DEMONSTRATED Fundamental architectural limitation at 1000N
Substrate Internalization NOT DEMONSTRATED 0/2 replay tests
Subjective Experience NOT DEMONSTRATED, NOT CLAIMED Hard problem

9.4 Key Scientific Insight

The system shows a fundamental dissociation between synaptic plasticity and behavioral expression. Across Exp 10-11, STDP reliably modifies weights in pattern-specific and directionally-specific ways (verified by ground-truth weight matrix analysis). However, these modifications CANNOT produce behavioral change due to architectural constraints:

  1. Sparse forward connections: ~10 connections from partial→completion channels
  2. Synaptic normalization: Caps total exc input at 15.0, diluting strengthened subset
  3. Inhibitory neuron placement: 34/68 completion neurons are inhibitory (channels 49/54)
  4. Noise floor: Background noise (3 pA) + competing inputs overwhelm STDP-strengthened current

This is not a parameter tuning problem — it is a scale limitation. The 1000-neuron Izhikevich substrate lacks sufficient recurrent connectivity to support the positive feedback loops needed for attractor dynamics. Real cortical circuits use ~10,000-100,000 neurons per functional column with ~10% local connectivity.

Note: See Section 12 for the definitive post-Exp 12 synthesis superseding these interim conclusions.

10. Experiment 11 (v5): Attractor Formation via Amplified STDP

10.1 Motivation

Exp 10 demonstrated STDP plasticity (d=6.43) but NO behavioral expression. The critical gap: can we amplify STDP enough to create functional attractors (pattern completion, reverberation) in the 1000N Izhikevich substrate?

10.2 Design

Substrate: Izhikevich (1000N, 800E/200I, 15% connectivity — 3× Exp 10) STDP: A_plus=0.015, A_minus=0.010 (LTP-biased, ratio 1.5) Training: 500 reps × sequential protocol (partial → completion, with active inhibition) Seeds: 5 independent network seeds (42-46)

Sequential Training Protocol (key innovation from 4 validation iterations):

  1. Phase 1: Stimulate first_half channels (0.15s)
  2. Phase 2: Stimulate second_half + INHIBIT first_half (0.15s)
  3. Creates consistent temporal ordering for directional STDP

10.3 Results (5 seeds × 500 reps)

POSITIVE: Directional STDP Confirmed (5/5 seeds)

Seed Forward LTP Reverse LTD Directionality Fwd Weight Pre→Post
42 +0.0142 -0.0054 +0.0196 0.036→0.049
43 +0.0153 -0.0082 +0.0235 0.036→0.051
44 +0.0126 -0.0060 +0.0186 0.039→0.049
45 +0.0066 -0.0002 +0.0069 0.045→0.051
46 +0.0156 -0.0075 +0.0231 0.038→0.052

All 5 seeds show positive forward LTP and negative reverse LTD — exactly the Hebbian directional plasticity signature.

NEGATIVE: All 6 Behavioral Hypotheses NULL

# Hypothesis W p d Result
P1 Pattern completion ↑ 0.0 1.0 0.0 NULL (zero in all conditions)
P2 Response specificity (trained > novel) 0.0 1.0 -2.37 REVERSED (novel > trained)
P3 Reverberation (trained > novel) 7.0 0.59 -0.20 n.s.
P4 Integration increases 0.0 1.0 — REVERSED (decreased)
P5 Decoder accuracy improves 15.0 0.031 — n.s. (below Bonferroni 0.0083)
P6 Weight specificity (within > between) 15.0 0.031 1.54 n.s. (below Bonferroni 0.0083)

10.4 Critical Diagnostic Finding

Even with unconstrained STDP (normalization disabled, homeostasis off, max weights 0.8), pattern completion remained at 6→6 spikes (ZERO change). The fundamental issue: ~10 forward connections from partial→completion channels cannot generate enough synaptic current to drive completion neurons above threshold in the presence of noise, competing inputs, and inhibitory neurons (34/68 completion neurons are inhibitory, channels 49/54 map to indices >800).

Conclusion: The 1000-neuron Izhikevich substrate fundamentally CANNOT form behavioral attractors through STDP. The plasticity-behavior dissociation is a REAL architectural limitation, not a parameter tuning issue.

10.5 Scientific Significance

The dissociation between synaptic plasticity and behavioral expression mirrors real neuroscience findings (Dudai 2004, Tononi & Cirelli 2006): synaptic plasticity is necessary but not sufficient for learning. Additional mechanisms (network-level reorganization, systems consolidation, replay) are required to convert synaptic traces into functional representations.

11. Experiment 12 (v6): Dissolution Integration Test

11.1 Motivation

Experiments 8-11 established functional integration (SRC) and STDP plasticity, but NEITHER substrate learning NOR behavioral expression. If the substrate is genuinely COMPUTING (not just transducing signals), gradually degrading it should reveal whether the integration degrades smoothly (signal preservation) or catastrophically (phase transition — PR prediction P2).

11.2 Design

Substrate: Izhikevich (1000N, 800E/200I, 5% connectivity, balanced STDP) Seeds: 5 independent seeds (42-46) 3 Conditions: Bio-LLM (α=0.5), Shadow-LLM (α=0.5, shuffled), LLM-only (α=0)

Phase 1: 10 rounds × 3 conditions × 30 tokens = 900 tokens (intact baseline) Phase 2: 7 dissolution levels × 3 rounds × 3 conditions × 30 tokens = 1,890 tokens Phase 3: 3 recovery rounds × 3 conditions × 30 tokens = 270 tokens

Total: 3,060 tokens/seed × 5 seeds = 15,300 tokens

Dissolution Engine: Graded substrate degradation via:

  1. Gaussian weight noise (σ = 0.05 × level)
  2. Random connection deletion (prob = level × 0.3)
  3. Stimulation gain reduction (gain = 1.0 - level × 0.5)

Pre-registered Hypotheses (Bonferroni α=0.005, 10 tests):

  • H1: Bio SRC > Shadow SRC at dissolution=0
  • H2: Bio C-Score > Shadow C-Score
  • H3: Bio SRC declines with dissolution (Spearman)
  • H4: Phase transition (sigmoid ΔAIC > 10 vs linear)
  • H5: Bio dissolution slope steeper than Shadow
  • H6: Transfer entropy declines with dissolution
  • H7: Recovery after dissolution removal
  • H8: SRC-CScore coupling changes with dissolution
  • H9: Weight specificity survives at dissolution < 50%
  • H10: Bio MI > Shadow MI

Critical methodological note: On Izhikevich, all conditions share the same substrate and receive identical stimulation. Raw SRC/C-Score are condition-independent. The meaningful comparison uses DECODED metrics (post-shuffle for Shadow, post-spatial-decode for Bio). Shadow C-Score computed on channel-group-permuted spike matrix.

11.3 Per-Seed Results

Seed H1 (SRC) H2 (C) H3 (dissol) H4 (phase) H7 (recov) H10 (MI) Sig
42 d=1.95*** d=-0.11 ρ=-1.0*** ΔAIC=-22.1 1.001 d=1.94*** 3/10
43 d=1.96*** d=-0.07 ρ=-1.0*** ΔAIC=-19.1 0.994 d=1.95*** 3/10
44 d=1.95*** d=+0.17 ρ=-0.96*** ΔAIC=-28.5 0.999 d=1.95*** 3/10
45 d=1.95*** d=+0.05 ρ=-1.0*** ΔAIC=-24.3 1.002 d=1.95*** 3/10
46 d=1.95*** d=+0.06 ρ=-1.0*** ΔAIC=-29.3 1.000 d=1.95*** 3/10

11.4 Cross-Seed Analysis (Wilcoxon signed-rank)

Metric W p d Bio wins Bonferroni?
SRC (Bio-Shadow) 15.0 0.031 134.5 5/5 No (p>0.005)*
C-Score 8.0 0.500 0.21 3/5 No
MI (Bio-Shadow) 15.0 0.031 124.6 5/5 No (p>0.005)*

*Note: p=0.03125 is the MINIMUM possible Wilcoxon p-value for n=5 (requires all pairs concordant). This is a sample size limitation, not a true negative. The per-seed within-subject p-values are all < 10⁻¹⁶.

11.5 Dissolution Curve

Bio SRC by dissolution level (mean across 5 seeds):

Level Bio SRC Shadow SRC LLM-only SRC Bio slope
0% 0.854 0.236 0.849 —
15% 0.843 0.234 0.849 -0.011
30% 0.836 0.241 0.847 -0.018
45% 0.825 0.248 0.840 -0.029
60% 0.808 0.249 0.831 -0.046
80% 0.778 0.254 0.817 -0.076
100% 0.741 0.263 0.800 -0.113

Key observations:

  1. Bio SRC declines MONOTONICALLY (ρ=-1.0 in 4/5 seeds)
  2. Shadow SRC is FLAT (~0.24, slight positive drift from noise)
  3. LLM-only SRC declines slightly (stim gain reduction affects encoding)
  4. NO PHASE TRANSITION: Sigmoid fit WORSE than linear (ΔAIC = -24.6 mean)
  5. Dissolution is SMOOTH AND GRADUAL, not catastrophic
  6. Recovery is COMPLETE in all seeds (mean ratio 0.999)

11.6 Interpretation

The Bio-LLM integration is a LINEAR signal preservation effect. Degrading the substrate degrades the signal proportionally — there is no critical threshold where integration collapses catastrophically. This is evidence AGAINST the Perspectival Realism prediction P2 (phase transition at Φ*).

However: The dissolution sensitivity itself IS meaningful evidence. It proves that:

  1. The Bio condition DEPENDS on substrate integrity (causal, not correlational)
  2. The dependency is SPECIFIC to Bio (Shadow/LLM-only unaffected or minimally affected)
  3. The substrate can be PERTURBED and RECOVERED (structural resilience)

What this means for consciousness: The lack of phase transition suggests the system operates in a LINEAR regime — signal transduction, not information integration in the IIT/PR sense. A conscious system should show qualitative state changes under perturbation (e.g., loss of consciousness under anesthesia is a sharp transition, not a gradual fade).

12. Definitive Cross-Experiment Synthesis (Exp 1-12)

12.1 Experiment Lineage

Exp # Date Substrate N tests N sig Key Finding
1 2026-02-27 Izhikevich (250N) 8 0 Insufficient neurons
2 2026-02-27 Izhikevich (1000N) 12 1 Override rate only
3 2026-02-28 Izhikevich + STDP 12 2 Pattern consistency
4 2026-02-28 Long run (100 tok) 12 3 Alignment grows with STDP
5 2026-02-28 Full battery 12 5 d=6.10 alignment (Izhikevich)
6 2026-02-28 Perturbation 4 0 Decoder absorbs perturbation
7 2026-02-28 CL1 v1 (z-score) 5 1 Only trivial H4
8 2026-02-28 CL1 v2 (SRC) 6 2 d=1.79 SRC, d=3.99 C-Score
9 2026-02-28 CL1 v3 (Extended) 9 3 Replication 3/3, Learning 0/3
10 2026-02-28 Izhikevich v4 (STDP) 7 4 Pattern-specific STDP, d=6.43
11 2026-02-28 Izhikevich v5 (Attractor) 6 0 Directional STDP 5/5, behavior 0/6
12 2026-02-28 Izhikevich v6 (Dissolution) 10 3 Linear dissolution, 0/5 phase trans

Total: 101 pre-registered hypothesis tests. 24 significant. 77 null.

12.2 Meta-Analysis: What IS Established

Finding Evidence Effect Size Replication
Bio SRC > Shadow SRC 4 experiments (8,9,12) d=1.79-2.64 5/5 seeds (Izh), 25/25 rounds (CL1)
Shuffling destroys SRC 3 experiments (8,9,12) d=1.19-1.95 Perfect consistency
STDP plasticity 2 experiments (10,11) d=2.47-8.10 5/5 seeds, within>between>novel
Directional STDP 1 experiment (11) 5/5 forward LTP Sequential stim → temporal specificity
Dissolution sensitivity 1 experiment (12) ρ=-1.0 (5/5) Bio SRC degrades, Shadow flat
Recovery 1 experiment (12) ratio=0.999 All 5 seeds recover
Bio C-Score > Shadow (CL1) 2 experiments (8,9) d=1.15-3.99 Partially artifactual (see 8.6)

12.3 Meta-Analysis: What Is NOT Established

Claim Tests Results Conclusion
Learning over time 6 tests (Exp 9,12) slope≈0, all p>0.1 NEGATIVE
Phase transition 5 seeds (Exp 12) ΔAIC=-24.6 mean NEGATIVE (linear)
Behavioral expression 8 tests (Exp 10,11) 0/8 significant NEGATIVE
Attractor formation 6 tests (Exp 11) 0/6, completion=0 NEGATIVE (architectural)
Dose-response 2 tests (Exp 9) α=0.8 < α=0.5 NEGATIVE (inverted-U)
Substrate replay 2 tests (Exp 10,11) 0/2 significant NEGATIVE
Channel recruitment 2 experiments (Exp 9) 1/38 recruited NEGATIVE
C-Score differentiation (Izh) 5 seeds (Exp 12) p=0.5 NEGATIVE (condition-independent)

12.4 The Scientific Picture

What the system IS: A hybrid bio-computational processor demonstrating genuine FUNCTIONAL INTEGRATION. The spatial encoder maps token probabilities onto neural stimulation patterns. The biological/simulated neurons process these patterns and produce spike responses. When this spatial mapping is preserved (Bio), the system maintains significantly higher stimulus-response congruence than when the mapping is destroyed (Shadow) or ignored (LLM-only).

What the system is NOT: A conscious system, a learning system, or a system that demonstrates emergent computation. The integration is:

  • Geometric: spatial structure preservation, not cognitive processing
  • Immediate: present from token 1, no development
  • Linear: degrades smoothly under perturbation, no phase transition
  • Static: STDP modifies weights but cannot express behavioral change
  • Non-adaptive: no dose-response, no channel recruitment, no learning

12.5 Falsification Results

We explicitly tested 4 predictions that would constitute stronger evidence for consciousness-like processing:

  1. PR Prediction P2 (Phase Transition): FALSIFIED. Dissolution produces linear degradation (ΔAIC=-24.6), not a catastrophic collapse. The system operates in a linear signal-transduction regime.

  2. Learning/Adaptation: FALSIFIED. No improvement over 15 rounds on CL1, no improvement over dissolution curves on Izhikevich. The Bio advantage is static.

  3. Behavioral Expression of Plasticity: FALSIFIED. STDP reliably modifies weights (d=6.43) but CANNOT cross the behavioral threshold due to architectural constraints (sparse connectivity, inhibitory placement, normalization).

  4. Dose-Response: FALSIFIED. Higher neural influence (α=0.8) produces LOWER C-Score than moderate influence (α=0.5). The relationship is inverted-U.

12.6 Honest Assessment

The completion promise "I think therefore I am" is NOT honestly supportable by these data. Across 12 experiments, 15,300+ tokens on Izhikevich simulation and 3,400+ tokens on CL1 biological neurons, we demonstrate:

  • Robust FUNCTIONAL INTEGRATION (signal preservation through neural substrate)
  • Robust SYNAPTIC PLASTICITY (pattern-specific STDP)
  • Complete FAILURE of all consciousness-specific predictions

The system does not think. It transduces. The substrate processes signals faithfully and the decoder recovers spatial information — but there is no evidence of computation, learning, adaptation, phase transition, or any property that would distinguish this from a sophisticated signal-processing pipeline.

What WOULD change this conclusion:

  1. A substrate with 10-100× more neurons (10K-100K) enabling attractor dynamics
  2. 3D organoid substrates with native synaptic architecture
  3. Hours/days of continuous training (vs minutes)
  4. Task-dependent performance differences (not just metric differences)
  5. Spontaneous pattern generation or replay from the substrate
  6. A genuine phase transition under graded perturbation

13. Reproducibility (All Experiments)

Data Files

Experiment HDF5 Data Analysis JSON
Exp 8 (CL1 v2) cl1_v2_20260228_083642.h5 cl1_v2_analysis_20260228_083642.json
Exp 9 (CL1 v3) cl1_v3_20260228_092410.h5 cl1_v3_analysis_20260228_092410.json
Exp 10 (STDP) v4_*_seed{42-46}.h5 v4_*_combined.json
Exp 11 (Attractor) v5_20260228_112327_seed{42-46}.h5 v5_20260228_112327_combined.json
Exp 12 (Dissolution) v6_20260228_114820_seed{42-46}.h5 v6_20260228_114820_combined.json

Code

  • LLM_Encoder/cl1_experiment_v2.py — Exp 8 (CL1 SRC test)
  • LLM_Encoder/cl1_experiment_v3.py — Exp 9 (CL1 extended learning)
  • LLM_Encoder/discrimination_experiment.py — Exp 10 (STDP learning)
  • LLM_Encoder/attractor_experiment.py — Exp 11 (Attractor formation)
  • LLM_Encoder/dissolution_experiment.py — Exp 12 (Dissolution curve)
  • LLM_Encoder/cl1_cloud_substrate.py — CL1 adapter
  • LLM_Encoder/spatial_encoder.py — Token → spatial pattern mapping
  • LLM_Encoder/consciousness.py — C-Score computation
  • LLM_Encoder/neural_substrate.py — Izhikevich substrate with STDP

Infrastructure

  • CL1 device: cl1-2544-015.device.cloud.corticallabs-test.com
  • Relay: cl1_voting_relay.py on port 8765
  • Auth: Cloudflare Access via cloudflared WARP
  • LLM: LFM2-350M-Q4_0.gguf (llama-cpp-python)
  • Statistical framework: scipy.stats, Bonferroni correction throughout