tostido/Butterfly-Field-Station-storage/work/Convergence_Engine/case_studies - Storage Bucket

Convergence Engine Case Studies

Language Composure, Understanding & Articulation

Budget: $1/hour total across 5 GPUs ($0.20/GPU/hour) Duration: 3-4 hours per experiment Focus: How organisms develop language comprehension and articulate responses

---

Research Questions

1. Composure: How do organisms learn to structure coherent responses?
2. Understanding: Can organisms demonstrate comprehension of input context?
3. Articulation: What conditions produce expressive, varied language?
4. Vocabulary Depth: Quality vs quantity of word usage
5. Semantic Precision: Using the RIGHT word, not just A word

---

Case Study Overview

GPU	Study	Key Manipulation	Research Question
1	Slow Learner	Low LR, high patience, small pop	Does slower learning = deeper understanding?
2	Immersion	Massive vocab exposure, no staged delay	Does vocabulary flood help or hurt composure?
3	Survival Pressure	High competition, fast battles	Does stress produce terse or rich language?
4	Social Learning	Large alliances, low competition	Does cooperation breed shared understanding?
5	Creative Specialist	High temperature, semantic guidance boost	Does creativity improve articulation?

---

Study 1: SLOW LEARNER

Hypothesis: Slower learning rate with longer training produces deeper semantic understanding

Key settings:

• learning_rate: 0.001 (5x slower than default)
• epsilon_decay: 0.995 (slower exploration decay)
• population_size: 7 (small, stable group)
• competition_intensity: 0.1 (minimal pressure)
• teaching_frequency: 1 (constant teaching)

Measure: Response coherence over time, vocabulary retention

---

Study 2: IMMERSION

Hypothesis: Immediate exposure to full vocabulary produces faster articulation

Key settings:

• staged_knowledge.enabled: false (no delay)
• vocab_size: 30000 (larger vocabulary)
• organism_embedding_alpha: 0.2 (faster embedding learning)
• temperature: 1.0 (neutral generation)

Measure: Unique words per response, time to first coherent output

---

Study 3: SURVIVAL PRESSURE

Hypothesis: High evolutionary pressure produces efficient, precise language

Key settings:

• competition_intensity: 0.5 (high)
• rounds_per_cycle: 3 (frequent battles)
• survival_threshold: 0.5 (harsh)
• population_size: 15 (more competition)
• germination_rate: 1.0 (fast replacement)

Measure: Word economy (meaning per word), response length vs quality

---

Study 4: SOCIAL LEARNING

Hypothesis: Allied organisms develop shared vocabulary and understanding

Key settings:

• max_alliance_size: 20 (large alliances)
• competition_intensity: 0.05 (very low)
• betrayal_chance: 0.0 (stable relationships)
• population_size: 25 (larger social group)
• phenotype_to_vocabulary: true (personality affects words)

Measure: Vocabulary overlap between allies, response similarity

---

Study 5: CREATIVE SPECIALIST

Hypothesis: Higher temperature + semantic guidance produces more expressive language

Key settings:

• temperature: 1.5 (more creative)
• semantic_boost: 0.4 (strong guidance)
• high_strength_boost: 0.2 (reward strong associations)
• concept_system.enabled: true (active concept formation)
• knowledge_web_influence_interval: 10 (frequent updates)

Measure: Vocabulary diversity, novel word combinations, articulation quality

---

Machine Setup

# On each GPU machine:
git clone <repo>
cd Convergence_Engine
pip install -r requirements.txt

# Copy the study config
cp case_studies/config_study_N.json config.json

# Fresh start
python clear_all_data.py

# Run (headless recommended)
python causation_web_ui.py

---

Data Collection

During Experiment (every 30 min)

1. Screenshot the organism chat panel
2. Note vocabulary counts in debug panel
3. Check Highlander battle count

After Experiment

1. data/logs/application.log - full system log
2. Chat with each surviving organism - save transcripts
3. Export atom formation data from debug panel
4. Note final population and vocabulary sizes

---

Analysis Metrics

Composure Score

• % responses without <UNK> tokens
• Average sentence length
• Grammar structure (subject-verb patterns)

Understanding Score

• Context relevance (does response match input?)
• Follow-up coherence (multi-turn consistency)
• Concept accuracy (uses concepts correctly)

Articulation Score

• Vocabulary diversity (unique words / total words)
• Semantic precision (word-concept alignment)
• Expression richness (adjectives, modifiers used)

Overall Language Quality

LQ = (Composure × 0.3) + (Understanding × 0.4) + (Articulation × 0.3)

---

Expected Results

Study	Composure	Understanding	Articulation
Slow Learner	HIGH	HIGH	MEDIUM
Immersion	LOW	MEDIUM	HIGH
Survival Pressure	HIGH	MEDIUM	LOW
Social Learning	MEDIUM	HIGH	MEDIUM
Creative Specialist	MEDIUM	LOW	HIGH

---

Post-Experiment Comparison

After all 5 complete:

1. Rank by overall Language Quality score
2. Identify which conditions produced best understanding
3. Check if any organism achieved multi-turn coherent dialogue
4. Compare vocabulary depth vs breadth across studies
5. Document emergent language patterns