A new benchmark for long-context language models is proposed. It measures a model's performance by its ability to correctly answer a series of 50 factual multiple-choice questions, where each question pertains to a single, isolated detail mentioned within a 100,000-word technical document. A model that answers all 50 questions correctly is deemed to have 'mastered' the document. Analyze the limitations of this evaluation approach. Specifically, explain why achieving a high score on this benchmark does not necessarily demonstrate a model's fundamental capability for comprehending the document in its entirety.