A Single-GPU Recipe for Specification-Grounded Industrial Decision Support

We study how to make a language-model agent reliable enough to support operators where a confident wrong answer is worse than none. Evaluated in three specification-bound industrial domains — thermal-power emissions control, water and wastewater treatment, and chemical-process safety, one instrumented as a deeper real-world case study — the assistant must answer from the site's procedures, use the right tool with the right arguments, and decline what it cannot ground. The safety-critical fields — alarm level, crossed limit, permitted action — come from a deterministic shield: a forecaster plus the written specification fixes them, and the agent relays them faithfully and acts around them. A single-GPU LoRA fine-tune on about 1,700 curated examples teaches the backbone to orchestrate that shield rather than substitute for it; under an identical toolbox the frontier also receives, it defers to the shield on the verbatim safety fields and outperforms on the rest. The binding result is one a third party can re-score: on a released benchmark scored deterministically against public gold, the agent leads a pinned open-weight baseline by +0.14 relay fidelity, +18 abstention, +18 safe refusal under adversarial pressure, and +18 stale-standard detection, conceding general domain QA by 1; the same ordering holds against a stronger hosted frontier given the identical toolbox, and the private case study corroborates under deeper instrumentation. Detecting a stale specification remains its primary open limitation, so it runs offline, on hardware each site already owns, under a human supervisor and a change-management process.