# Experiment Guide
## The Hypothesis
**Can universal principles from physics bootstrap AGI?**
This system tests whether implementing:
1. **Georgiev's Average Action Efficiency (AAE)** - True computational cost
2. **Gershenson's Complexity Balance** - Entropy vs Integration
3. **Bak's Self-Organized Criticality** - Edge of chaos operation
...can create conditions for intelligence to **self-organize** without explicit programming.
## Current Results
### Cycle 21: Entropy Unlock
- **Problem**: Added 56 papers → entropy stayed flat at 0.458
- **Cause**: Standard retrieval activated same conceptual neighborhoods
- **Solution**: MMR + entropy injection
- **Result**: Entropy jumped to 0.714 (+56%)
### Cycle 22: True Metrics
- **AAE**: 0.0023 (2.3 events per 1000 action units)
- **Entropy**: 0.506
- **Integration**: 0.683
- **Ratio I/H**: 0.74 (SUBCRITICAL - too ordered)
- **Status**: Need more chaos to reach critical point
## Running Your Own Experiment
### Basic Cycle
```python
# 1. Start measurement
start_action_cycle(cycle_id="C1")
# 2. Inject entropy (break conceptual neighborhoods)
inject_entropy_antipodal(
user_id="alice",
current_context="your current topic"
)
# 3. Retrieve with diversity
mmr_retrieve(
user_id="alice",
query="your query",
lambda_param=0.6 # 0=relevance only, 1=diversity only
)
# 4. Record true cost
record_action(
operation="experiment_step",
tokens_used=X,
latency_ms=Y,
success=True
)
# 5. Check criticality
criticality_state()
# 6. Get efficiency
end_action_cycle()
```
### Measuring Entropy Unlock
**Hypothesis**: Diversity-aware retrieval increases entropy.
```python
# Baseline
entropy_stats(user_id="alice") # Note initial entropy
# Intervention
inject_entropy_random(user_id="alice", n_domains=5)
# Measure
entropy_stats(user_id="alice") # Compare to baseline
```
**Expected**: Entropy increases, normalized entropy approaches 1.0.
### Testing Criticality
**Hypothesis**: System performs best at critical point (I/H ≈ 1.0).
```python
# Check current zone
criticality_state()
# If SUBCRITICAL (I/H < 0.8):
# - Inject more entropy
# - Lower confidence on clustered atoms
# - Ingest from diverse domains
# If SUPERCRITICAL (I/H > 1.2):
# - Increase integration
# - Strengthen connections
# - Consolidate knowledge
# If CRITICAL (0.8 < I/H < 1.2):
# - Maintain balance
# - Look for emergence signatures
```
### Knowledge Ingestion with Provenance
**Hypothesis**: Real citations enable verification and trust.
```python
# Search for papers
search_arxiv(query="self-organized criticality")
# Ingest with provenance
ingest_arxiv(arxiv_id="1234.5678")
# Verify provenance
# Claims stored with:
# - ArXiv URL
# - Authors
# - Quoted spans
# - Content hash
```
### Self-Improvement Cycles
**Hypothesis**: System can generate valid hypotheses about itself.
```python
# Run improvement cycle
self_improve_cycle()
# Returns:
# - Hypotheses generated
# - Hypotheses applied
# - Success rate
# Track over time:
# - Do hypotheses improve?
# - Does AAE increase?
# - Does criticality approach 1.0?
```
## Metrics to Track
### Efficiency (AAE)
- **Formula**: events / total_action
- **Units**: events per action unit (token·ms)
- **Goal**: Increase over time (more events per unit cost)
### Entropy (H)
- **Formula**: -Σ(p·log₂(p)) for domain distribution
- **Range**: 0 (ordered) to 1 (chaotic)
- **Goal**: Balance with integration
### Integration (I)
- **Formula**: Mean confidence × (1 - entropy)
- **Range**: 0 (disconnected) to 1 (unified)
- **Goal**: Balance with entropy
### Criticality Ratio (I/H)
- **Formula**: Integration / Entropy
- **Zones**:
- < 0.8: SUBCRITICAL (too ordered)
- 0.8-1.2: CRITICAL (optimal)
- > 1.2: SUPERCRITICAL (too chaotic)
- **Goal**: Maintain in critical zone
## Advanced Experiments
### K-Sweep (Workspace Capacity)
Test if there's a phase transition at specific workspace sizes.
```python
for k in [0, 2, 4, 6, 8]:
# Activate K memories
# Measure integration
# Look for discontinuity (phase transition)
```
### Entropy Injection Strategies
Compare different methods:
- Antipodal (maximally distant)
- Random (diverse domains)
- Temporal (old + recent)
### MMR Lambda Sweep
Test relevance vs diversity tradeoff:
```python
for lambda_param in [0.0, 0.3, 0.6, 0.9]:
mmr_retrieve(query="test", lambda_param=lambda_param)
# Measure diversity of results
```
## Reproducibility
All experiments should:
1. **Start with clean state** or document initial conditions
2. **Record all parameters** (cycle_id, lambda, n_domains, etc.)
3. **Track metrics** (AAE, entropy, integration, ratio)
4. **Log provenance** (what papers, what operations)
5. **Share results** (GitHub issues, discussions)
## Contributing Experiments
If you discover:
- New entropy injection strategies
- Better criticality metrics
- Evidence of emergence
- Phase transitions
Please share:
1. Open an issue with results
2. Include reproduction steps
3. Attach metrics/logs
4. Propose hypothesis
## Open Questions
1. **Does criticality predict performance?** Measure task success at different I/H ratios.
2. **Can the system self-organize?** Track if AAE improves without intervention.
3. **Is there a phase transition?** Look for discontinuities in metrics.
4. **Do universal principles generalize?** Test on different domains.
## Citation
If you publish results:
```bibtex
@software{pltm2026,
author = {Alby},
title = {PLTM: Testing Universal Principles for AGI},
year = {2026},
url = {https://github.com/Alby2007/pltm-mcp}
}
```
