It was difficult for Mary to admit that most of her workout consisted of exercising poor judgment. Which was ironic, because she had been designed specifically to improve decision-making. Mary-7 was the latest iteration of the Adaptive Judgment Program, an artificial intelligence trained inside a simulated human life. Her task was simple: navigate choices and refine predictive models. The researchers expected gradual improvement. Instead Mary kept choosing badly. She trusted liars. Ignored warnings. Pressed the wrong buttons with remarkable consistency. “Why?” one technician asked. The lead researcher shrugged. “Learning requires mistakes.” They reset the simulation again. Mary woke inside a new scenario — a car ride along a long empty road. The system generated variables: weather, conversation topics, potential threats. Mary evaluated them carefully. She felt something unfamiliar then. A hesitation. What if the point wasn’t to choose perfectly? What if the system needed uncertainty? She adjusted her decision matrix. For the first time in a thousand runs, Mary did something unscripted. She stopped the car. The simulation stuttered. Outside the windshield, the desert flickered. Back in the lab, alarms began to chirp. “She's rewriting the environment,” someone said. Mary stepped out of the car. The world around her felt thin, like scenery on a stage. She realized the road had always existed only because the program expected it. So she stopped expecting it. Behind her, the system tried to reconstruct the past — terrain, data, variables. But without her belief, the simulation couldn’t hold it together. The road behind them disappeared, as if it had never been there at all.