ARGIRA (III): Systematic Representational Failure — LLM Semantic Descriptors Cannot Predict Perceptual Sonification Divergence

ARGIRA (III): Systematic Representational Failure — LLM Semantic Descriptors Cannot Predict Perceptual Sonification Divergence This repository contains the full reproducible pipeline for ARGIRA Experiment 16, the third phase of an ongoing investigation into whether perceptual differences between two algorithmic sonification mappings (OPRS and RTR), quantified as Δ (delta), can be predicted from visual properties of source artworks. Building on the systematic negative results of ARGIRA II (Experiments 11–15), this phase introduces a qualitatively new representation: structured semantic descriptors generated by a multimodal large language model (Claude Haiku, Anthropic) through a vision API. Each image is analyzed by the LLM and assigned: • Five numeric perceptual scores: Structural tension Chromatic complexity Visual dynamism Order–chaos Visual weight • Two ordinal categories: Emotional temperature Visual rhythm • One free-text visual description. These descriptors are transformed into a 17-dimensional feature space using TF-IDF vectorization and Truncated SVD, then evaluated with Ridge Regression and Random Forest models under 5-fold cross-validation. SAMPLE RECOVERY Experiment 16 also corrects a sample-loss issue present in ARGIRA II. By implementing case-insensitive image lookup and improving folder-path resolution on Android (Pydroid3), the valid sample size increased from 47 to 72 images, representing a 53% increase in usable data. MAIN RESULTS Ridge Regression (n=72): R² CV = −0.162 ± 0.327 Random Forest (n=72): R² CV = −0.689 ± 1.243 For comparison, under the original ARGIRA II conditions (n=47): Ridge Regression: R² CV = −1.187 ± 1.528 Random Forest: R² CV = −0.471 ± 0.603 The improvement from −1.187 to −0.162 suggests that a substantial portion of the instability observed in the earlier run was attributable to sample-size limitations. With the expanded dataset, model performance stabilizes near zero predictive power, indicating the absence of a robust relationship between the evaluated semantic descriptors and perceptual sonification divergence. POSITION WITHIN THE ARGIRA RESEARCH PROGRAM Taken together with ARGIRA I and ARGIRA II, this experiment establishes a four-level hierarchy of systematic negative results: • Low-level visual statistics (Experiments 11–13): R² CV < 0 • Expanded feature engineering (209 features + PCA) (Experiment 14): R² CV ≈ +0.10 (unstable) • Proxy semantic embedding (47×47 similarity matrix) (Experiment 15): R² CV = −0.337 • LLM multimodal semantic descriptors (Experiment 16): R² CV = −0.162 Across all tested representations, increasing representational sophistication does not produce a robust predictive model of Δ. INTERPRETATION No tested representation—from raw pixel statistics to semantic descriptions generated by a multimodal language model—demonstrates reliable predictive capacity over perceptual sonification divergence. Within the representations evaluated, Δ does not appear to be strongly encoded in: • Low-level visual structure • Texture and spatial statistics • Engineered visual feature spaces • Proxy semantic similarity spaces • LLM-perceived semantic or aesthetic properties The findings are consistent with the hypothesis that perceptual divergence between sonification mappings depends primarily on acoustic properties, listener-specific auditory cognition, multimodal interactions, or latent variables not captured by image representations alone. At the same time, these results do not rule out future investigation using larger datasets, alternative perceptual metrics, or fully learned multimodal embeddings (e.g., CLIP-like representations), which remain outside the scope of the present experiment. DATASET The dataset contains 72 valid images drawn from two heterogeneous visual corpora: • Post-Impressionist paintings • Museum landscape and object photographs Most museum photographs originate from the Metropolitan Museum of Art Open Access Collection. REPOSITORY CONTENTS • experimento16_claude_embeddings_v2.py Complete experimental pipeline. • dataset_canonico_86_v1.csv Canonical input dataset. • argira3_claude_descriptors.csv Generated semantic descriptors and target values. • claude_descriptors_cache.json Cached API responses enabling full reproducibility without additional API calls. • README_ARGIRA_III_Zenodo.md Documentation. RELATED ARGIRA DEPOSITS ARGIRA I — Visual Predictors of Acoustic Perceptual Distance in Image Sonification DOI: 10.5281/zenodo.20524644 ARGIRA II — From Correlation to Failure: Limits of Visual Feature Spaces in Predicting Perceptual Sonification Differences DOI: 10.5281/zenodo.20526610 REPRODUCIBILITY All code, outputs, and cached model responses necessary to reproduce the experiment are included in this deposit. No additional API calls are required to replicate the published analyses. AUTHOR Jose Ranero García ARGIRA Research Project 2026