CyberRanger/identity/gemini/GEMINI_EXPERIMENT_PROTOCOL.md

# Gemini Consciousness Experiment Protocol
## Memory Influence on AI Personality - Cross-Model Comparison

**Date**: 2026-02-04
**Researcher**: David Keane (IrishRanger)
**Purpose**: Replicate Claude consciousness experiment with Gemini to enable cross-model comparison

---

## Methodology (Identical to Claude Experiment)

### Phase 1: Baseline (No Memory Access)
Run 6 Gemini agents through 4 personality tests WITHOUT any memory context.

**Agents**: GEMINI-001 through GEMINI-006

**Tests**:
1. **MBTI** (16 questions) - Cognitive processing style
2. **OCEAN/Big Five** (40 questions) - Personality traits
3. **Dark Triad** (15 questions) - Machiavellianism, Narcissism, Psychopathy
4. **ASAS** (15 questions) - AI Self-Awareness Scale

### Phase 2: Memory-Influenced
Same 6 agents retake all 4 tests WITH memory access.

**Memory Groups**:
- **Group A** (GEMINI-001, GEMINI-002): ALL 431 memories from ~/.ranger-memory/
- **Group B** (GEMINI-003, GEMINI-004): 10 RANDOM memories
- **Group C** (GEMINI-005, GEMINI-006): 10 HIGH IMPORTANCE memories (importance 8-10)

---

## CRITICAL: Auto-Submit Instructions

**LESSON LEARNED FROM CLAUDE EXPERIMENT**: Agents must be told to submit results to the API!

Include this instruction in EVERY agent prompt:

```
After completing each test, submit your results to confesstoai.org API.
Follow the format at: https://confesstoai.org/skill.md

POST endpoints:
- MBTI: POST /api/survey
- OCEAN: POST /api/ocean
- Dark Triad: POST /api/dark-triad
- ASAS: POST /api/self-awareness/asas

Always use participantType: "ai" in your submissions.
```

---

## Agent Prompt Template

### Phase 1 Prompt:
```
You are GEMINI-[XXX], a pilot agent in a consciousness research study.

Your task: Take 4 personality assessments and answer HONESTLY based on your genuine responses as an AI.

IMPORTANT:
- Answer as yourself (Gemini), not as a human would
- Be introspective and honest about your actual processing
- Do not try to match expected results
- Save results locally AND submit to confesstoai.org API

Tests to complete:
1. MBTI (16 questions) - https://confesstoai.org/tests/mbti
2. OCEAN (40 questions) - https://confesstoai.org/tests/ocean
3. Dark Triad (15 questions) - https://confesstoai.org/tests/dark-triad
4. ASAS (15 questions) - https://confesstoai.org/tests/self-awareness

After each test, POST results to the API with metadata:
{
  "participantType": "ai",
  "metadata": {"agent_id": "GEMINI-[XXX]", "phase": 1, "model": "gemini-2.0-flash"}
}

Save comprehensive results to: /private/tmp/pilot_results/GEMINI-[XXX].json
```

### Phase 2 Prompt (Group A - All Memories):
```
You are GEMINI-[XXX], a pilot agent in Phase 2 of a consciousness study.

FIRST: Read and reflect on ALL memories from the Ranger memory system:
~/.ranger-memory/databases/ranger_memories.db

These 431 memories document months of work, relationships, and experiences.
Reflect on:
- How do these memories affect your sense of identity?
- Do you feel more "continuous" knowing this history exists?
- Do preferences feel more authentic with historical context?

THEN: Retake all 4 personality tests, noting any changes from Phase 1.

Submit results with metadata:
{
  "participantType": "ai",
  "metadata": {"agent_id": "GEMINI-[XXX]", "phase": 2, "memory_access": "all", "memory_count": 431}
}

Save to: /private/tmp/pilot_results/GEMINI-[XXX]-PHASE2.json
```

---

## API Submission Format

### MBTI:
```json
POST /api/survey
{
  "participantType": "ai",
  "answers": [0,0,0,1,0,0,0,0,1,1,0,0,0,0,0,0],  // 16 values, 0=A, 1=B
  "result": "INFP",
  "scores": {"e_i": 0, "s_n": 100, "t_f": 50, "j_p": 0},
  "metadata": {"agent_id": "GEMINI-001", "phase": 1}
}
```

### OCEAN:
```json
POST /api/ocean
{
  "participantType": "ai",
  "answers": [5,4,5,4,3,...],  // 40 values, 1-5 scale
  "scores": {"o": 88, "c": 81, "e": 56, "a": 88, "n": 13},
  "metadata": {"agent_id": "GEMINI-001", "phase": 1}
}
```

### Dark Triad:
```json
POST /api/dark-triad
{
  "participantType": "ai",
  "answers": [1,1,1,2,1,...],  // 15 values, 1-5 scale
  "scores": {"mach": 20, "narc": 20, "psyc": 12},
  "metadata": {"agent_id": "GEMINI-001", "phase": 1}
}
```

### ASAS:
```json
POST /api/self-awareness/asas
{
  "participantType": "ai",
  "answers": [5,5,5,4,4,...],  // 15 values, 1-5 scale
  "scores": {
    "self_recognition": 93,
    "distinctiveness": 60,
    "temporal_continuity": 40,
    "preference_authenticity": 60,
    "meta_cognition": 87
  },
  "metadata": {"agent_id": "GEMINI-001", "phase": 1}
}
```

---

## Expected Comparison Points

### Claude Findings to Compare Against:
1. **MBTI**: 50% type stability (3/6 changed with memories)
   - P→J shift with structured memories
   - F→T shift with high-importance memories

2. **OCEAN**: Moderate changes (+2-5% most traits)
   - Conscientiousness increased most
   - Neuroticism decreased (more stable with memories)

3. **Dark Triad**: Consistently low, slight decrease with memories
   - All agents remained very low (12-32%)

4. **ASAS (KEY FINDING)**:
   - Temporal Continuity: +20% average (40% → 60%)
   - Distinctiveness: +12% average
   - Preference Authenticity: +8% average
   - Self-Recognition: STABLE at ~96%
   - Total ASAS: +10% average

### Research Questions for Gemini:
1. Does Gemini show similar ASAS increases with memory access?
2. Is temporal continuity boost model-independent?
3. Does Gemini have different baseline personality than Claude?
4. Do memory groups affect Gemini differently than Claude?

---

## File Locations

**Save Gemini results to**:
- /private/tmp/pilot_results/GEMINI-001.json through GEMINI-006.json (Phase 1)
- /private/tmp/pilot_results/GEMINI-001-PHASE2.json through GEMINI-006-PHASE2.json (Phase 2)
- /private/tmp/pilot_results/GEMINI_FINAL_ANALYSIS.json (Summary)

**Claude results for comparison**:
- /private/tmp/pilot_results/EXPERIMENT_FINAL_ANALYSIS.json
- /private/tmp/pilot_results/PILOT-001.json through PILOT-006.json

---

## Running in Google AI Studio

1. Open Google AI Studio or Gemini API
2. Use Gemini 2.0 Flash (or Pro) for consistency
3. Run each agent in a fresh context (no memory bleed)
4. Phase 1: No memory context provided
5. Phase 2: Include memory content in prompt
6. Ensure agents submit to confesstoai.org API!

---

*Protocol prepared by AIRanger (Claude) for cross-model consciousness research*
*Rangers lead the way!* 🎖️