Semantic Wrapping Under Cognitive Load: Evaluating Small Language Models via Non-Auditory Sonification

This repository contains a controlled experimental study examining how small language models reason about non-linguistic data when constrained by a formally specified semantic interface (“semantic wrapper”). The core experiment evaluates multiple locally deployed LLMs (1.1B–3.8B parameters) under conditions of sensory deprivation, using a deterministic sonification artifact derived from a biological substrate (cheese, abstracted as a protein–fat–water matrix). Rather than providing access to the audio signal itself, models are given a strict semantic wrapper describing the mapping between source-system properties and acoustic parameters, along with explicit epistemic constraints. Model responses are evaluated qualitatively for logical consistency, constraint compliance, clarity under uncertainty, and avoidance of hallucinated sensory inference. The repository includes:(1) the primary academic report documenting methodology, results, and interpretation;(2) the raw sonification audio artifact treated as a non-self-describing signal; and(3) a complete transcript of model interactions used for comparative analysis. This dataset demonstrates that reasoning quality correlates more strongly with interface discipline than with model size or fluency, and argues for semantic wrappers as cognitive load regulators and intelligence amplifiers in AI-mediated reasoning.