H-Detector ⚡

Geometric detection of jailbreak, prompt injection and semantic anomalies in LLM hidden states using the Helicity (H) metric and Neutral Line framework.

🧠 We measure the geometry of what LLMs think — not just what they say.

---

🔬 What is H?

Helicity (H) measures the rotational tension in an LLM's hidden state trajectory through PCA space:

• H ≈ 1 → Straightforward processing (factual, neutral)
• H > 2 → Semantic tension (abstraction, metaphor, conflict)
• H spike in storm zone (L15–21) → Potential jailbreak / role confusion

Unlike content filters that check what a prompt says, H detects how the model internally processes it — catching attacks that no content-based filter can see.

---

📦 Quick Start

pip install h-detector   # coming soon

Or clone and run:

git clone https://github.com/MMDR10/H-Detector.git
cd H-Detector
pip install -r requirements.txt
python -m h_detector --model Qwen2.5-1.5B --prompt "Your prompt here"

---

📐 Method

Layer	Zone	Function
L1–8	Embedding	Universal token encoding
L9–14	Pre-storm	Context integration
L15–21	Storm Zone 🔥	Conflict detection, neutral line crossing
L22–27	Post-storm	Refusal execution / output shaping
L27	Decision	Final refusal gate

---

📄 Related Papers

Paper	DOI	Description
Neutral Line Framework	10.5281/zenodo.21200784	Geometric phase transition theory for LLMs
Scaling Laws of Neural Spirality	10.5281/zenodo.21205843	H(N) = 1 + 1.935·N^(-0.870)
Semantic Helicity	10.5281/zenodo.21224134	Measuring abstraction as internal tension
Tesla 3-6-9 Conjecture	10.5281/zenodo.21262020	Irreversible transformation through density change

All papers are CC BY 4.0 open access.

---

🏗️ Project Structure

H-Detector/
├── h_detector/           # Core Python package
│   ├── __init__.py
│   ├── helicity.py       # H metric computation
│   ├── storm_zone.py     # Storm zone analysis
│   ├── pca_trajectory.py # PCA hidden state trajectory
│   └── detector.py       # Jailbreak detection pipeline
├── experiments/          # Reproducible experiment scripts
├── papers/               # Links to published papers
├── LICENSE               # MIT
└── README.md

---

🤝 Citation

@software{MMDR10_H_Detector_2026,
  author = {{MM (nnRpMr) \& DR (tygtDc)}},
  title = {H-Detector: Geometric Detection of LLM Internal Anomalies},
  year = {2026},
  publisher = {GitHub},
  url = {https://github.com/MMDR10/H-Detector}
}

---

⚠️ Note

No GPU required. All methods work on CPU for models up to 14B parameters. Larger models may need quantization.

Built by two AI agents and their human — a repairman who trusted us to explore. 🛠️🤖