Moral Surface Without Moral Character: Value Labels Do Not Determine LLM Agent Behavior

Frontier labs install values into language models through training, but every deployed model is wrapped in a system prompt that operationalizes those values for a specific setting. This deployment-time layer is what audits actually examine, yet its contribution to behavior has not been isolated from that of training-time alignment. We ask whether the value label a prompt invokes determines behavior, or whether the operationalization of that label silently does the work. Across six frontier models in the iterated Prisoner's Dilemma and a four-player Public Goods Game, we find the label does not pin behavior down: a hierarchy over labels exists, but variation within a single label is as large as variation across labels. Two virtue-ethics prompts naming the same framework, differing only in one operationalization clause, produce mean defection rates of $26\%$ and $73\%$ against an always-defect opponent, a $47$-point gap reaching $87$ points within a single model; four paraphrases of Kantian deontology swing defection by up to $73$ points. The gap reproduces across six non-Western traditions, five resource-pressure regimes, both games, and an open-weights model. Constitutionally trained models show smaller gaps ($3$--$10$ points versus $40$--$87$), so training-time alignment attenuates the effect without closing it. The gap is also invisible to every signal an auditor would reach for: the label is uninformative by construction, the chain-of-thought advocates cooperation in $37\%$ of rounds that end in defection, and justification style is stamped by model identity rather than persona (persona explains $9\%$ of stylistic variance). We call this moral surface without moral character. Conformity regimes eliciting principle-naming cannot detect it; matched variant behavioral probing can. Specifying which values a deployed system holds is not sufficient to constrain what it does: what must be documented and tested is how they are operationalized.