Constraint Obedience, Silence Tolerance, and Semantic Reframing in Sub-5B Language Models

This document presents a qualitative diagnostic study of locally hosted language models, focusing on behavioral performance under explicit constraints rather than benchmark accuracy or task success. Using five minimal, reproducible prompts, the study evaluates constraint obedience, silence tolerance, and semantic reframing strategies across multiple model architectures and parameter scales. All models were tested under identical conditions (single-turn prompts, fresh context, local inference, no system priming). Behavioral outputs are analyzed qualitatively to identify recurring failure modes and training-dependent strengths. The results indicate that training philosophy—particularly perceptual grounding in vision-language models—plays a stronger role than parameter count in enabling semantic inhibition and pragmatic silence. No claims are made regarding general intelligence or benchmark superiority. The findings are limited to observable behavioral patterns under controlled diagnostic conditions.