RMA-JPEG v1: Auditing Structural Information Preservation Under JPEG Compression

RMA-JPEG is a case study within the Residual Model Auditor (RMA) framework, a statistical auditing methodology designed to measure how much structural information survives a deterministic transformation.In this release, the audited transformation is JPEG compression: the framework evaluates whether measurable properties of the original image can still be recovered from the compressed representation. Unlike conventional image-quality benchmarks centered on perceptual similarity, RMA-JPEG approaches compression as an information-preservation problem. The objective is not merely to ask whether compressed images “look similar,” but whether statistically meaningful descriptors of the source image remain inferable after compression. Structural Preservation Under JPEG Compression (n = 109 real paintings, JPEG quality factor = 10) This is the initial public release of RMA-JPEG. Although conceptually related to the broader ARGIRA ↔ RMA research line (concept DOI: https://doi.org/10.5281/zenodo.21131137), this package is not a continuation or version extension of that series.RMA-JPEG uses a distinct Layer 2/3 statistical engine based on: correlation analysis, mutual information, PCA decomposition, feature importance ranking, and bootstrap stability estimation, rather than the Layer 1 causal auditing engine (rma_core_v3.py) used in the ARGIRA sonification framework. The previous works are therefore referenced strictly as related conceptual foundations, not as prior versions of this package. Central claim On a fixed corpus of 109 real paintings, JPEG compression at quality factor = 10 preserves chromatic statistics near-perfectly (r ≥ 0.98 for all color descriptors, r = 1.000 for luminance), while simultaneously introducing measurable and statistically stable degradation in edge-based structural metrics — most notably in the primary edge-density metric (r = 0.77, 95% CI [0.73, 0.82], bootstrap n = 500). Additionally, a methodological artifact present in an earlier adaptive-threshold formulation of edge density was identified, isolated, and excluded from all official analyses reported in this release. Explicit scope This release is intentionally narrow in scope. The claims reported here apply only to: a corpus of 109 real paintings, compressed with JPEG quality factor = 10, evaluated through the specific descriptors implemented in this release. No generalization is claimed regarding: JPEG compression as a whole, other quality levels, other image domains, perceptual quality, or semantic reconstruction fidelity. Those questions remain future work. Explicitly out of scope for v1 The following domains are intentionally excluded from this release: RMA-OCR, which belongs to a different epistemological domain (symbolic reconstruction distortion rather than compression auditing), Candidate 3 (Texture Structure), Candidate 5 (Compression Artifacts), synthetic stress-testing datasets, and cross-domain generalization experiments. These are reserved for future releases or separate publications. Package contents release_v1/ ├── README.md │ Figure → metric → claim mapping │ ├── RMA-JPEG_RESULTS_CONSOLIDATED.md │ Canonical results table (correlations, PCA, feature importance) │ ├── VALIDATION_LOG.md │ Reimplementation-based validation vs. legacy (all deltas = 0.0) │ ├── CHANGELOG.md ├── VERSION.txt ├── LICENSE │ Dual license: MIT (code) + CC BY 4.0 (data/docs/figures) │ ├── rma_core/ │ Shared Layer 2/3 statistical engine │ ├── __init__.py │ ├── correlation.py │ ├── mutual_information.py │ ├── pca.py │ ├── feature_importance.py │ ├── bootstrap.py │ ├── reporting.py │ ├── pipeline.py │ │ Public API: run_layer2() │ └── utils.py │ ├── RMA-JPEG_package/ │ ├── corpus/ │ │ ├── candidate1_corpus.csv │ │ │ Color statistics, n=109 │ │ └── candidate2_corpus.csv │ │ Edge structure, n=109 │ │ │ ├── scripts/ │ │ └── build_candidate4_analysis.py │ │ │ └── results/candidate4_results/ │ ├── candidate4_global.csv │ │ Merged corpus (109 × 33) │ ├── candidate4_correlations.csv │ │ 32×32 correlation matrix │ ├── candidate4_mutual_information.csv │ ├── candidate4_pca.csv │ │ Explained variance per component │ ├── candidate4_feature_importance.csv │ │ Random Forest importances (256 rows) │ ├── candidate4_bootstrap_summary.csv │ │ Mean + 95% CI per variable │ ├── candidate4_bootstrap_pc1.csv │ └── candidate4_summary.md │ ├── figures/ │ ├── fig1_correlation_bars.png │ │ Pearson r with CI, all C1+C2 variables │ ├── fig2_scatter_comparison.png │ │ GT vs JPEG scatter (luminance vs edge density) │ ├── fig3_bootstrap_distribution.png │ │ Bootstrap histogram, edge_density_fixed │ ├── fig4_pca_scree.png │ │ Variance explained per PC │ └── fig5_artifact_comparison.png │ Fixed vs. adaptive threshold artifact │ └── supplementary/ └── candidate4_bootstrap.csv 500 raw bootstrap iterations (full CI reproducibility) Reproducibility All numerical outputs included in this release were verified through direct comparison between the modular rma_core implementation and the corresponding legacy outputs used during earlier development stages. Every validated file produced a numerical delta equal to 0.0. This validation process is intentionally described as: reimplementation-based validation against legacy outputs and not as validation of a “historical refactor,” because an earlier narrated refactor phase was later discovered to have no corresponding executable code artifacts.The distinction is documented explicitly in VALIDATION_LOG.md. The confidence interval associated with every reported correlation is fully reproducible from: supplementary/candidate4_bootstrap.csv which contains all 500 raw bootstrap iterations used in the analysis. Equivalent results can also be regenerated programmatically through: rma_core.bootstrap.bootstrap_analysis(random_state=42) License This repository uses a dual-license structure: MIT License → source code (rma_core/, analysis scripts) CC BY 4.0 → documentation, result CSVs, figures, and supplementary material The Zenodo metadata license is set to CC BY 4.0 for repository-level compatibility. Related identifiers ARGIRA ↔ RMA (sonification case study)Concept DOI: https://doi.org/10.5281/zenodo.21131137 Residual Mapping Audit (RMA) v3 core — Layer 1 causal auditingConcept DOI: https://doi.org/10.5281/zenodo.21048837 How to cite Ranero García, Jose. (2026).RMA-JPEG v1: Structural Preservation Under JPEG Compression at Quality Factor = 10 (n = 109).Zenodo. https://doi.org/10.5281/zenodo.21157733