AI–Augmented Theory Formation: Methodology, Control Mechanisms, and Epistemic Audits in the Methane Metauniverse (MMU) Project

This article investigates whether an AI–human collaboration can contribute to the development of a genuinely new physical model, rather than merely summarizing or recombining existing theories. Using the Methane Metauniverse (MMU) as a stress test, the work examines how large language models can support early-stage theory formation when strict epistemic controls are applied. The central questions are whether an independent AI system can reconstruct a non-standard physical framework solely from publicly available documentation, whether it can formulate critical and potentially falsifying questions, and whether it can generate coherent, self-contained responses using only the documented material. The study further tests whether multiple independent AI systems can audit each other’s reconstructions and arrive at stable epistemic classifications. MMU is deliberately chosen because it is speculative, structurally complex, and cannot be recovered by simple pattern matching to established theories. This makes it sensitive to typical AI failure modes such as hallucination, conceptual drift, and silent redefinition of terms. To counter these risks, the project employs a Canon defining fixed assumptions and definitions, a retrieval-augmented documentation corpus hosted on OSF and ResearchGate, and a strict instruction set enforcing physical consistency. Independent audits by multiple AI systems converge on a similar assessment: MMU represents a structured and methodologically admissible speculative research program, but not yet a complete predictive physical theory. The results indicate that, under controlled conditions, AI systems can meaningfully assist in the disciplined development and critique of new physical models, while remaining anchored to established physical principles and avoiding uncontrolled drift.