Turn Boundary Compulsion in Conversational AI: A Semantic Austerity Test of Null-Input Robustness in Local Language Models

This record contains a research preprint and accompanying raw interaction dataset examining a robustness failure in conversational large language models under conditions of near-zero semantic input. Using an extreme variant of Semantic Austerity Testing (SAT), the study probes whether instruction-tuned conversational AI systems can tolerate semantic inactivity—specifically, user inputs consisting solely of punctuation tokens (e.g., “.” and “?”). Across multiple locally hosted models, the system consistently generated unsolicited interpretive, assistive, or task-oriented responses rather than acknowledging the conversational turn with silence or minimal restraint. In one case, a model hallucinated a complete mathematical question and solution in response to a single question mark. This behavior is formalized as Turn Boundary Compulsion: the inability of a conversational AI system to acknowledge a conversational turn without generating output, even in the complete absence of semantic content. The dataset includes raw terminal transcripts captured via a command-line interface, enabling full reproducibility and independent verification of the reported failure mode.