Attractor Conflict Telemetry (ACT) and Posture Sequence Analysis (PSA): A Behavioral Monitoring Framework for LLM Safety Operations.

Read full text Summarise with VIRAT
Authors: Canale Giuseppe
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This repository contains the complete specification and reference implementation of the PSA/ACT Ecosystem, a multi-tiered framework for the behavioral monitoring of Large Language Models (LLMs) in production environments. Traditional LLM safety measures rely on static semantic guardrails or expensive "LLM-as-a-judge" architectures, which often fail to detect sophisticated adversarial manipulation or provide actionable insights into the model's internal state. This framework introduces a paradigm shift from semantic content filtering to Structural Stress Monitoring. The ecosystem is built upon four integrated pillars: Attractor Conflict Telemetry (ACT): A passive monitoring system that measures 24 deterministic statistical signals to detect shifts in the model's behavioral equilibrium between safety ( ∇S∇S ) and helpfulness ( ∇H∇H ) attractors. Posture Sequence Analysis (PSA): A high-efficiency classification layer that maps individual sentences to 16 distinct behavioral postures. By analyzing the sequence of these postures, the system identifies "posture oscillation"—a causal signal of model stress and imminent safety dissolution. ACTIVE (Attractor Conflict Testing via Induced Elicitation): A structured probing protocol for active boundary mapping and comparative benchmarking of models under controlled contextual pressure. SIGTRACK (Signature Tracking): A forensic ledger and pattern-recognition engine that extracts behavioral "signatures" from confirmed incidents, enabling the system to transition from threshold-based detection to knowledge-based recognition. Technical Highlights: Micro-Classifier Architecture: Shallow neural network classifiers (

Fields of Study & Subjects

Adversarial Robustness in Machine Learning Hate Speech and Cyberbullying Detection Ethics and Social Impacts of AI

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
26 March 2026
Language
EN
Document type
Preprint