Concepts

Veridex is built on a few core concepts that differentiate it from other detection libraries.

The Signal Architecture

At the heart of Veridex is the Signal. A Signal is an atomic unit of detection logic. It takes a specific type of input (e.g., text string, image array) and outputs a DetectionResult.

Why Signals?

• Independence: Each signal runs in isolation. If one fails (e.g., due to missing dependencies), it doesn't crash the pipeline.
• Explainability: You can inspect the output of each signal individually to understand why content was flagged.
• Extensibility: Adding a new detection method is as simple as subclassing BaseSignal and implementing the run() method.

Probabilistic Scoring

Unlike binary classifiers that output "Fake" or "Real", Veridex signals output a Score and a Confidence.

• Score: A value between 0.0 (Human) and 1.0 (AI).
• Confidence: A value between 0.0 and 1.0 indicating how reliable the signal considers its own assessment.

Fusion (Planned)

While individual signals are useful, they often have weaknesses. We are working on Fusion strategies to aggregate multiple signals into a final verdict. This will allow combining weak signals (like perplexity) with strong signals (like DetectGPT) to produce a robust final decision.

(Fusion implementation is currently in development.)