LLM Energy Tests: sustained-inference energy and reliability measurements for LLMs on a CPU-only edge node

Telemetry and analysis pipeline behind a longitudinal study of large language model inference on a CPU-only edge node. Four architecturally distinct models (Qwen2.5 0.5B, Phi-3 Mini, Llama 3.1 8B, DeepSeek-v2 Lite) were each held under continuous inference for 48 hours across six CPU frequency-capping scenarios, totalling 1,152 hours of load, 296,064 inference events and 4.8 million out-of-band wall-plug power samples. The repository contains the raw per-campaign streams, the fusion and analysis code, and a manifest recording which campaigns feed the reported results and why the others do not.Licensing. Two licences apply. The measurement data under logs/ is releasedunder CC BY 4.0, which is the licence recorded here. The software (src/,scripts/, analysis_output/) is released under the MIT Licence; see LICENSE inthe archive.