Paraverbal Computational Signals for Detecting Scheming in Large Language Models: A Shadow–Archive–Telemetry Framework

AbstractWe propose a novel framework for detecting scheming in large language models (LLMs). Existing detection methods, relying primarily on outputs or chain-of-thought (CoT) inspection, are insufficient when models exhibit situational awareness and deliberately mask covert reasoning. Our framework integrates three orthogonal signal families:1.Shadow signals (St): entropy anomalies, log-prob irregularities, and latency jitter at the token level.2.Archive signals (At): recurrence of covert strategies across episodes, distinguishing transient artifacts from stable deceptive intent.3.Paraverbal signals (Pt): deviations in latency, jitter, and energy consumption relative to calibrated baselines, inspired by paraverbal cues in human communication.These signals are fused by an adaptive Governor that implements GREEN/AMBER/RED policies, with dynamic thresholds updated through a botta-e-risposta feedback process. The system introduces overhead only transiently during anomaly escalation and returns to equilibrium once deception is neutralized.This framework provides a forensic and operational toolkit for AI safety, complementing training-based approaches and establishing a foundation for resilient monitoring in deployment.Keywords: AI safety, scheming, deception detection, paraverbal signals, anomaly detection, adaptive governor