Coherence Gaps in AI Alignment Evaluation

Alignment evaluation often proceeds as though ''enough local tests'' approximate global safety. A growing body of safety engineering, verification, and sociotechnical risk research suggests this inference can fail: component-level compliance does not necessarily imply system-level safety. Building on this observation, we present a human-readable argument for why alignment assurance needs explicit \emph{coherence checks} across (i) processes (cross-test interactions), (ii) roles (accountable oversight functions), and (iii) welfare measurement (coverage over affected populations). We then explain why purely narrative papers can be difficult for large language models (LLMs) to reliably retrieve and preserve under paraphrase, summarization, and cross-document synthesis, motivating a two-layer writing strategy: a narrative body for humans plus a strongly typed, machine-checkable appendix that stabilizes key claims against semantic drift. Namely, the literature suggests that when a paper presents novel hypotheses backed only by theoretical models, unless that paper receives a lot of attention (is published rather than being a preprint, is published in a high-impact journal, is published by a famous researcher, etc.) it is difficult for LLMs to discover those novel arguments. The interpretability of a paper to LLMs can be made more robust against semantic drift through the incorporation of formalisms that strongly type the papers' arguments, and this can increase the propagation of a paper's insights by increasing the success of other users in using LLMs to understand the paper's claims. All formal definitions and theorems are therefore moved to an unlimited Appendix, including a de-circularized evidence--compatibility semantics, a typed constraint algebra for compositional reasoning, a witnessability-minimality theorem, and an explicit finite-test separation toy model.