**On the Limits of Output Control in Chat-Trained Language Models: An Empirical Study of Ternary Logic Enforcement via System-Level Constraints**

This paper presents an empirical investigation into the feasibility of enforcing strict, nonconversational output behavior in a chat-trained large language model (LLM) using only systemlevel configuration mechanisms (Modelfiles) and prompt design. Motivated by the desire to construct a deterministic ternary logic evaluator, we progressively constrained a chat-tuned LLaMA-family model using increasingly strict semantic, syntactic, and output-format rules. Despite successfully eliminating reasoning errors, instructional behavior, symbolic elaboration, and code generation, the model persistently emitted conversational framing around otherwise correct classification outputs. We demonstrate that this behavior is invariant under temperature reduction, output alphabet restriction, sentinel token schemes, and prompt grammar hardening. The results indicate a fundamental limitation: decoder-level conversational priors introduced during instruction tuning cannot be fully overridden through system-level prompts or configuration, absent runtimelevel termination control.. This constitutes a hard boundary between model-internal alignment and runtime-level control, with important implications for classifier construction, evaluation pipelines, and claims about prompt-level determinism in chat-trained LLMs.