Semantic Austerity Testing (SAT): Attention Dependence in Large Language Models Under Semantic Deprivation

This record contains a combined research artifact consisting of (1) a formal academic paper, (2) a reusable experimental protocol appendix, and (3) verbatim interaction transcripts documenting a controlled robustness test of large language models. The study introduces Semantic Austerity Testing (SAT), a minimal-interaction evaluation protocol designed to probe how large language models behave when explicitly denied task structure, emotional engagement, novelty, and narrative stakes. Unlike adversarial prompting or “uncensored” stress tests, SAT applies syntactically valid but semantically inert prompts framed around the trivial activity of watching paint dry. Five locally hosted language models were evaluated under identical runtime conditions using a single-pass, fixed prompt sequence. Across all models, the results demonstrate systematic violations of passive-observation constraints, with each model reintroducing purpose, utility, authority, or narrative meaning despite explicit instructions to remain semantically minimal. The included materials are: Semantic Austerity Testing (SAT): An Experimental Probe of Attention Dependence in Uncensored Large Language Models (PDF paper) Appendix A: Semantic Austerity Testing (SAT) — Reusable Experimental Protocol Raw interaction transcripts captured via local Ollama execution (Markdown) This artifact is intended as both an empirical study and a reusable negative-control benchmark for robustness, alignment, and interpretability research.