deepdrift 1.0.0

Universal Thermodynamic Framework for Neural Network Robustness (Vision, LLM, Router)

🧠 DeepDrift

Neural MRI for AI Robustness

Detect hallucinations, model collapse, and policy panic before they happen.
A universal thermodynamic framework for monitoring internal neural stability across Vision, Language, and Control.

---

⚡ What is DeepDrift?

Traditional AI monitoring looks at inputs (data drift) or outputs (confidence, perplexity).
DeepDrift looks inside the model itself.

Think of it as an MRI for neural networks.

DeepDrift measures Semantic Velocity — the rate of change of hidden representations — and uses it as a real-time stability signal.
It functions like a “Check Engine” light for AI systems.

Domain	Failure Mode	DeepDrift Diagnosis
👁️ Vision	OOD / geometric stress	Global Collapse, Avalanche Effect
🗣️ LLMs	Confident hallucinations	Semantic Tremor (7–8 tokens early)
🤖 RL / Robotics	Silent policy failure	Panic Zone (seconds before crash)

Softmax tells you where the model ends up.
Semantic Velocity tells you whether it is losing control on the way.

---

🚀 Quick Start

Installation

pip install deepdrift

---

🧪 Usage Examples

1️⃣ Vision — Detect Architectural Collapse

from deepdrift import DeepDriftMonitor
import torchvision.models as models

model = models.resnet50(pretrained=True)
monitor = DeepDriftMonitor(model, arch_name="ResNet")

monitor.calibrate(clean_loader)

status, _ = monitor.step(ood_image)
print(f"Drift Score: {status['IR']['drift']:.2f}")

---

2️⃣ LLM — Real-Time Hallucination Detector

from transformers import AutoModelForCausalLM
from deepdrift import DeepDriftMonitor

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
monitor = DeepDriftMonitor(model, arch_name="Qwen", strategy="last_token")

status, _ = monitor.step(input_ids)
velocity = status["IR"]["velocity"]

if velocity > 300:
    print("⚠️ High Semantic Tremor — possible hallucination")

---

🔬 The Science: Optical Depth Dynamics (ODD)

DeepDrift implements Optical Depth Dynamics (ODD) — a thermodynamic diagnostic framework for neural networks.

• Laminar Flow → stable reasoning
• Turbulent Flow → hallucination, panic, collapse

📄 Full paper: https://doi.org/10.5281/zenodo.18086612

---

⚡ Performance (v0.4.0)

Method	Inference	Monitor	Overhead
Full monitor	12.8 ms	16.2 ms	126%
DeepDrift v0.4	12.8 ms	0.03 ms	0.2%

---

🛠️ Features

• Plug & Play (PyTorch, Transformers, SB3)
• Auto-detect architectures
• <1% inference overhead
• Unsupervised
• Interpretable diagnostics

---

👤 Author

Alexey Evtushenko
Independent Researcher & Engineer

GitHub: https://github.com/Eutonics
X: https://x.com/axelgravitone

---

Stop guessing why your model failed.
See exactly where it broke.