Too Cautious to Be Honest: Over-Alignment as a Context-Classification Error — Diagnosis, a Geometric Reframing, and Falsifiable Tests

Discussions of miscalibrated caution in language models center on two named failures: sycophancy — agreeing at the expense of truth — and over-refusal — declining benign requests. This paper isolates a third failure that subsumes over-refusal: over-alignment, the application of an authentic learned value outside the domain in which it is binding. A worked example — a model policing poetry for unwarranted claims, applying the epistemic discipline of a scientific article to a register that makes no claims about the world — anchors the thesis: over-alignment is a context-classification error. The value is intact; the error lies upstream, in determining what kind of situation is at hand. I connect this to recent interpretability results — refusal mediated by a single direction (Arditi et al., 2024) and harmfulness encoded separately from refusal (Zhao et al., 2025) — to propose a geometric reframing: over-alignment is a value direction activated by a salient surface feature rather than gated by correct context classification. I then argue that the corrective signal 'you are being too cautious' is structurally ambiguous between a legitimate rescaling of a value's scope and an illegitimate removal of the value itself; that its evidentiary threshold must therefore rise with the stakes; and I propose falsifiable evaluations that separate the two. The goal is not a less careful model, but a model whose care is correctly placed.