Telemetry Degradation and Sequential Stability in LLM-Integrated IoT Decision Systems
TL;DR
This work introduces a framework to evaluate the behavioral stability of LLM decision systems transforming IoT telemetry to semantic output, and argues behavioral stability in LLM-integrated IoT systems must be evaluated jointly with sensing reliability and regime structure.
Abstract
AI systems for the human performance/interface must transform raw Internet of Things (IoT) telemetry into consistent actionable output. Large language models (LLMs) function as such interpreters. However, behavioral stability through real-world data degradation remains unexplored; applications demanding consistent output (e.g., medicine and athletics) remain unrealized. We introduce a framework to evaluate the behavioral stability of LLM decision systems transforming IoT telemetry to semantic output. The framework is validated in a case study of AI-tennis coaching using stroke-level IoT smartwatch streams. Local perturbations are mapped to global output through graded dropout, resolution downsampling, and reliability signaling. We establish the baseline temporal stability under idealized conditions and then quantify the impact of increasing degradation severity on the sequential decision behavior. Temporal volatility (FlipRate), directional escalation, and cumulative safety outcomes are assessed. We show telemetry degradation significantly increases decision instability in already unstable systems. In a balanced regime, FlipRate rises from 0.144 under clean conditions to 0.207 under downsampling (+44%), while in a serve-dominant regime it reaches 0.300. Regimes characterized by strong feature dominance exhibit amplified sensitivity to information loss. A workload accumulation model shows observed volatility constitutes a structural pathway to safety-bound violations. Reliability signaling alone produces smaller effects than structural information loss, indicating informational cues cannot compensate for lost telemetry. Structured prompting partially mitigates instability but does not eliminate degradation-induced effects. We argue behavioral stability in LLM-integrated IoT systems must be evaluated jointly with sensing reliability and regime structure. Our robustness analysis provides a basis for LLM-telemetry interpretation, opening new AI-implementation avenues in human performance and digital health.