Supernetwork Architecture Selection for Deployable Language Models: RAG Assessment of Overhead-Line Insulators on a 2 GB Device for Grid Operator's Maintenance Crews

Transmission system operators inspecting high-voltage insulator strings using inference using drone capture inference in remote locations need a local LLM assistant that is useful only if its answers are hallucination-free, drawn from operator-held documents and never exceeding what a photograph shows. The assistant presented here is built accordingly, a detector turning photographs into fixed-text findings, retrieval covering 3,780 IEC, CIGRE and IEEE passages and 127 engineer-reviewed question-answer pairs, and code-enforced hallucination guards checking every answer. Because everything runs within the memory restrictions of an edge device such as the 2 GB Arduino UNO Q, the language model must fit the device, so a weight-sharing supernetwork is trained once from Llama 3.2 1B and eleven subnetworks varying in size are extracted and distilled from a Qwen3.5-27B teacher, yielding a deployable model family. Which network is selected is decided by a selection rule, evaluation set and GPT-OSS:20B judge, ranking candidates by answer quality under a hallucination guard gate and memory restrictions. All subnetworks keep guard violations at or below 2.8% and answer the field questions on the device without failure at 2.8 W, the service peaking at 1108 to 1258 MB as quality rises along the memory-quality Pareto frontier from 0.73 to 0.91 and the median 704-token answer lengthens from 79 to 134 s. The shipped network is the fastest of the three tied on quality within 1200 MB, at 0.85, 94 s per answer and 1172 MB.