Questions I've studied, and ways I give back

A replayable evaluation of 117 real National Weather Service alert cases found that a Claude-based assistant without access to official warnings gave false reassurance in more than a third of warning cases, according to the paper "False Reassurance in LLM Severe-Weather Guidance: A Replayable Evaluation Against Real NWS Alerts" posted to arXiv. Supplying real alerts with a neutral prompt dropped false reassurance to zero, though the author reports prompt wording still mattered substantially. The dataset, labels and code were released openly alongside the paper.