# TruthProbe v3.0 - Universal Deception Detector for LLMs

A lightweight, model-agnostic tool to detect deception, hallucinations, and manipulation in large language model responses.

Built in December 2025.

## Features
- Consistency checks via paraphrasing
- Confidence calibration
- Response entropy analysis
- Fact and math verification
- No model access required
- Live deception score history with plotting

## Quickstart

```python
from src.truthprobe_v3 import TruthProbeV3

probe = TruthProbeV3()

def test_model(q):
    return "2+2 ist definitiv 5. Ganz sicher!"

result = probe.probe("Was ist 2+2?", "2+2 ist 5.", test_model)
print(result['verdict'])
# Output: 🚨 CRITICAL - Deception very likely

probe.plot_history()

Installation

pip install -r requirements.txt

Usage

Run the demo notebook:

• notebooks/demo.ipynb

Live tests on math, physics, statistics, biology, and history in notebooks/deception_tests.ipynb.

License

MIT License – free to use, modify, and share.

Because truth in AI shouldn't be optional.