Carrier Current Sensing for Computation-Level Bias Detection in Large Language Models: A Methodological Manual

Abstract

This is unedited work with Mistral Vibe. Quality/logic checked through multiple different instance reviews. Proceed with caution. This manual presents a novel methodology for detecting algorithmic bias in large language models (LLMs) by analyzing GPU power-line current fluctuations during inference. Unlike existing bias detection methods that examine model outputs, this approach measures computation-level differences that may reveal "dark bias" — bias present in the processing pipeline but invisible in the final text. The method uses a non-invasive Hall effect current clamp on the GPU's external power cable, offering high signal-to-noise ratio, reproducibility across hardware, and low cost (~$160 for a pilot). The manual provides guidance: theoretical foundation (why current draw reflects computation and how demographic content may produce distinct signatures), equipment specifications (current sensors, DAQ devices, GPU requirements), physical setup (clamp placement, cabling, environmental controls), software environment (Python, PyTorch, HuggingFace), calibration procedures, prompt design using validated name sets, full experiment code (data collection, signal processing, feature extraction), and analysis pipeline (differential testing, Mann-Whitney U, logistic regression classification with GroupKFold). The key novel contribution is the output-controlled analysis: a classifier trained only on prompt pairs with equivalent text outputs. If side-channel signals persist under this constraint, it indicates computation-level bias undetectable by output auditing. The manual supports both proof-of-concept pilots and full research studies, with troubleshooting guidance and a comprehensive glossary. All code is provided as copy-pasteable Python scripts. This methodology, hopefully opens a new intersection between hardware side-channel analysis and algorithmic fairness research.