A physicist turned ML engineer argues that large language models shouldn't decide when a forecast shifts. Modern NWP systems like the ECMWF IFS deliver 9 km forecasts every few hours, yet developers often fetch the data, feed it to an LLM, and ask “Did the weather change?” The author's chaos‑theory training warns that deterministic thresholds are lost in probabilistic prompts.

He built Skygent, a hybrid agent that isolates the decision step in pure Python. A Pydantic‑validated evaluator compares consecutive forecast snapshots, applies configurable deltas for temperature, precipitation or wind, and respects skill limits that vary with horizon. Only when a threshold is crossed does the pipeline invoke GPT‑4o‑mini to turn the structured JSON into a short, human‑readable alert.

In practice the system polls every six hours, triggering alerts roughly once or twice weekly per event, versus dozens of LLM calls for a naïve design. At $0.0001 per narrative the cost is negligible, and each decision is traceable through explicit variable, delta and confidence fields. The method fails when boundaries cannot be codified, but keeping thresholds out of prompts yields reproducible, explainable alerts.