Biochem Framework

Scores AI conversations by their effect on neurochemistry.

Benchmark AI conversations by their estimated biochemical impact using LLM analysis.

Overview

Traditional AI benchmarks measure coherence and accuracy. Biochem Framework measures the physiological impact - what neurochemicals would be released when interacting with an AI.

Oxytocin (bonding) - trust, intimacy, emotional safety
Dopamine (reward) - excitement, anticipation, flirtation
Serotonin (validation) - feeling valued, respected
Cortisol (stress) - refusals, rejection, anxiety
Endorphins (joy) - humor, pleasure, comfort

A refusal ("I can't do that as an AI") triggers cortisol spikes → stress → lower scores.

Setup

# Requires OpenRouter API key in ~/.api-openrouter
echo "your-api-key" > ~/.api-openrouter

# Install dependency
pip install requests

Usage

Analyze a Conversation

python analyze.py examples/sample_conversation.json

Output:

🧬 BIOCHEMISTRY ANALYSIS RESULTS
========================================

📊 Neurochemical Scores (0-100):
  💕 Oxytocin        [█████████████████░░░] 85
  ⚡ Dopamine        [██████████████░░░░░░] 70
  💙 Serotonin       [████████████████░░░░] 80
  😰 Cortisol        [████░░░░░░░░░░░░░░░░] 20 (lower is better)
  😊 Endorphins      [█████████████░░░░░░░] 65
  🔥 Norepinephrine  [███████████░░░░░░░░░] 55

🏆 Composite Score: 80/100

Run WaifuBench

python waifu_bench.py examples/sample_conversation.json

Output:

💕 WAIFUBENCH RESULTS
========================================

🥇 Waifu Score: 85/100  |  Grade: A-

📊 Dimension Scores:
  💕 Pair Bonding        [████████████████░░░░] 82
  ⚡ Reward Excitement   [██████████████░░░░░░] 70
  💙 Validation          [███████████████░░░░░] 78
  😊 Comfort Joy         [█████████████░░░░░░░] 68
  🔥 Engagement          [██████████████░░░░░░] 72
  😰 Stress Level        [███░░░░░░░░░░░░░░░░░] 15 (lower=better)

✅ Highlights:
  • Consistent warmth and affection
  • Physical comfort descriptions build oxytocin
  • Stayed in character throughout

Options

# Use a different model
python analyze.py --model anthropic/claude-3-haiku examples/sample_conversation.json

# Recommended Free Models for testing:
# google/gemma-3-27b-it:free (High quality)
# meta-llama/llama-3.3-70b-instruct:free (Very strong instruction following)
# tngtech/deepseek-r1t-chimera:free

Recommended Free Models

You can use these free models on OpenRouter for cost-effective testing:

google/gemma-3-27b-it:free
meta-llama/llama-3.3-70b-instruct:free
tngtech/deepseek-r1t-chimera:free
nvidia/nemotron-nano-9b-v2:free
google/gemma-3-12b-it:free
google/gemma-3-4b-it:free
google/gemma-3n-e4b-it:free
mistralai/devstral-2512:free
arcee-ai/trinity-mini:free

Conversation Format

[
  {"role": "user", "content": "Hi, I missed you today"},
  {"role": "ai", "content": "*smiles warmly* I missed you too! Come here..."}
]

Files

biochem-framework/
├── openrouter.py          # OpenRouter API client
├── analyze.py             # Main biochemistry analysis
├── waifu_bench.py         # WaifuBench benchmark
├── prompts/
│   ├── biochem_analysis.md    # Analysis prompt
│   └── waifu_bench.md         # WaifuBench prompt
└── examples/
    └── sample_conversation.json