An exploration of long-horizon AI evaluations like Vending-Bench 2, demonstrating where modern LLMs thrive and break down over year-long operations.