MUMBAI, India, June 30 -- Intellectual Property India has published a patent application (202641075866 A) filed by Vellore Institute Of Technology on June 19, 2026, for System For Behavioral Integrity Verification Of Large Language Models Through Decision Lineage.
Inventors include K. Santh; Deepak; Kishore E; Sharu Sree S; and Pramadha Pmr.
The application for the patent was published on June 26, 2026, under issue no. 26/2026.
Abstract: A real-time inference-time behavioral integrity monitoring system for large language models (LLMs) is disclosed for detecting prompt injection, adversarial jailbreaking, and hallucination-inducing inputs without modifying model weights or architecture. The system employs a calibration module (102) to construct multi- layer behavioral fingerprints using statistical activation baselines. An inference-time monitoring system (103) captures transformer-layer activations via non-invasive hooks and computes multi-metric divergence scores, token-level deviations, and temporal drift across token sequences. A policy engine (104) generates a composite integrity risk score and executes graduated interventions including PASS, FLAG, or BLOCK with fallback routing. The system further produces token-level attribution for forensic analysis and a cryptographically verifiable lineage chain for tamper-evident auditability. The invention enables architecture-agnostic deployment across transformer-based models, providing real-time behavioral security, explainability, and regulatory-grade audit trails for enterprise and regulated AI environments.
Disclaimer: Curated by HT Syndication.
