Minimax 2.7: If the model updates itself, who is responsible for the behavior of the version running in production?

M2.7 introduces a new complexity: a model that improves during deployment is also a model whose behavior can change over time. How can we ensure the reproducibility of the output? Regarding evaluation: if the model self-optimizes, the benchmark you used to choose it could change... AIs in 2026 are evolving, not static, systems. They optimize themselves, and in certain contexts, they modify themselves. Concrete examples of agentic AI in enterprises demonstrate this: governance, auditing, and accountability for these systems require processes that most organizations haven't yet built. How do you think this will impact open source? Maybe I don't quite understand their technology. What do you think?