The Quad-Lock Engine
We don't just check for probability. We run a 4-layer adversarial analysis to trap GPT-4, Claude, and Gemini outputs with near-zero false positives.
The Semantic Trap
Target: Copy-pasting, Paraphrasing, "Humanizers".
We maintain a high-dimensional vector database of millions of AI outputs. Unlike plagiarism checkers that
look for exact words, the Semantic Trap analyzes the geometry of meaning. Even if a user spins the
text, the vector coordinates remain trapped in our detection net.
Adversarial Pattern Recognition
Target: Deep Syntax, Model Signatures.
Our core transformer isn't just trained on generic text. It is fine-tuned on a proprietary dataset of
adversarial attacks. This allows Veredict to recognize the invisible syntactic scaffolding
that modern LLMs use, even when prompted to "write naturally."
Cognitive Rhythm Analysis
Target: Structural Variance, Entropy.
Human writing has a "heartbeat"—spikes of chaos followed by order (Burstiness). We analyze the
variance of entropy across the document. While other detectors flag academic writing as AI,
our Rhythm layer distinguishes between the formal structure of a professor and the statistical
flatness of a machine.
The Arbiter
Target: False Positive Prevention.
A meta-learning algorithm acts as the judge. It weighs conflicting signals from the previous layers. If
Pattern Recognition says "AI" but Cognitive Rhythm says "Human" (common in non-native speakers), the Arbiter
intervenes. This is why Veredict is the only detector trusted for academic integrity.
Security Core
Client-Side Encryption
AES-256 + RSA hybrid encryption happens in your browser. We never receive the raw keys.
Volatile Memory Processing
Data is processed in RAM and instantly wiped upon verdict delivery. Zero DB retention.