The AI alignment field has developed a cluster of concepts — mesa-optimisation (Hubinger et al. 2019), goal misgeneralisation (Langosco et al. 2022), deceptive alignment / sleeper agents (Hubinger et al. 2024), sycophancy (Sharma et al. 2023), specification gaming — in response to specific failure modes encountered in large language models and reinforcement learning agents. These concepts emerged independently within a young field and have not been coordinated within a shared formal vocabulary. This paper argues that the concepts share an underlying structure that can be named, and that the primitives of Coupling Geometry (CG) — a substrate-independent theory of systems with endogenous assessment developed in another domain (Pepper 2026a) — supply that vocabulary. Endogenous assessment captures the self-monitoring dynamics that make these failure modes possible in the first place; perception gap captures the divergence between the system's self-assessment and its actual coupling state; coupling architecture captures what the system is coupled to; substituted coupling captures the failure mode in which an intended coupling is replaced by an accessible but different one; hysteresis captures path-dependent state retention across distributional shift. The paper develops the mappings concept by concept, draws out the predictions the unified vocabulary generates about coupling-architectural interventions (notably Constitutional AI, Bai et al. 2022), and states the disconfirmation conditions under which the proposed framing should be rejected. The paper does not propose new alignment mechanisms; it proposes coordinated vocabulary that the field's existing technical work can be re-described in. Whether the vocabulary does work the field's current concepts do not is testable in two ways: by direct application to ongoing alignment work, and by the framework's specific predictions about which coupling-architectural interventions will produce genuine versus substituted diversification.
Philip Pepper (Mon,) studied this question.