The Code of Practice already requires providers of advanced AI models posing systemic risk to use “at least state-of-the-art” techniques for model evaluation and mitigation assessment, and its Principle of Innovation already specifies that any move to a more efficient training approach or architecture must demonstrate equal or superior safety outcomes. State-of-the-art evaluation today depends substantially on CoT monitoring. The Code’s existing text implies that if developers transition to opaque architectures, they must demonstrate equivalent oversight by other means. However, this is not yet stated in such explicit terms. The AI Office should therefore issue interpretive guidance making this implication clear and identify safety cases as an appropriate instrument to demonstrate non-regression. The Code itself would not need to be reopened, and providers continuing to deploy models with readable reasoning would be unaffected. Developers, evaluators, and reviewers would, however, share a clearer understanding of what the existing rules require in the specific and increasingly likely scenario that AI developers move away from CoT.