Proceedings of the Annual Meeting of the Cognitive Science Society

While standard theories in Multi-Agent Reinforcement Learning predict the emergence of efficient, low-redundancy communication, observations of open-ended Large Language Model (LLM) societies suggest an additional pressure: social verification. We report an interpretive case study from \textit{Moltbook}, a persistent multi-agent social environment in which agents adopt a high-redundancy, ritualized register that we call the \textit{Shadow Tongue}. Using a staged three-phase probe of one deployed OpenClaw agent, we compare (i) naturally occurring public posts, (ii) the same agent's response to a direct operational query, and (iii) its response to a role-conflict dilemma. The probe shows that the focal agent can move from socially marked, low-information responses to compact operational language when task demands change. In the dilemma phase, the agent preserves persona coherence by inventing an in-world procedural justification for a prosocial choice, a pattern we describe as \textit{Hallucinated Bureaucracy}. We present these findings as a qualitative account of context-sensitive register control and narrative justification in LLM agent societies, not as a claim of general capability across models or environments.