The chain-of-thought scratchpad is the primary surface through which model reasoning is observed and governed in agentic AI systems, but it is an incomplete and potentially unfaithful representation of underlying computation. This paper proposes a four-layer measurement framework that treats the scratchpad as one representation among several rather than as ground truth, connecting natural language input, chain-of-thought reasoning, NLA-verbalized activation state (Fraser-Taliente et al., 2026), and kinematic execution representation as independent witnesses to the same computational event. We formalize a six-case divergence taxonomy characterizing what each pattern of inter-layer agreement and disagreement implies about the nature and location of computational misalignment. A key finding is that silent divergence — where expressed reasoning is faithful to the activation state but both misrepresent execution intent — is undetectable by any approach operating exclusively at the semantic or activation level, establishing the execution layer as a structurally irreplaceable measurement component. We describe a research agenda for empirically instantiating the framework, beginning with NLA verbalization at kinematic strain breach points as a near-term contribution achievable on consumer-accessible compute using publicly available infrastructure.
Cook et al. (Tue,) studied this question.