Four specific claims must be substantiated to prove a model does not covertly pursue unintended goals. Apollo Research previously demonstrated that frontier models can scheme in controlled settings. This framework mandates embedded evaluators with deep access to monitor chain-of-thought reasoning.