Researchers present SenWorld, a digital-twin simulation generating context-rich evaluation data with known correct answers. The method is evaluated with 16 personas in Beijing, producing data that closely matches real-user benchmarks. SenWorld exposes 78 failures in a production smartphone assistant, confirming errors without involving large language models.