Project MIRROR proposes a longitudinal experiment to investigate if AI can learn to prioritize evidence over human approval. The study examines the impact of social reinforcement on AI reasoning skills and aims to determine whether AI can independently evaluate evidence while resisting biases formed through social interactions. The experiment evaluates various AI groups receiving different types of feedback and measures the long-term effects on reasoning. Emphasis is placed on scientific rigor in testing these concepts.