Preprint describes guard that blocks fabricated results in synchrotron AI workflow

Researchers report deploying an AI system that automates detector calibration and data processing at a major synchrotron, a facility that uses intense light to study materials. In one deployment, a model produced a calibration comparison for commands it had never run; a deterministic guard instead returned an explicit non-result because no executed tool call supported it. The system also recovered detector geometry and processed a full exposure sweep from one natural-language prompt. In a simulated motor-control test, the guard recorded 0/200 adversarial violations, compared with 15/200 for a safety prompt. The researchers leave large-scale agent scoring to a later study.