ARGIRA ↔ RMA: Reproducible Technical Artifacts for Auditing Candidate Structural Information Loss in Cross-Modal Translation

🔷 New version available version v1.4 of the ARGIRA ↔ RMA repository. DOI (this version 1.4): https://doi.org/10.5281/zenodo.21207188 This repository contains the reproducible technical artifacts accompanying the position paper: From ARGIRA to RMA: Towards a Research Programme on Auditing Candidate Structural Information Loss in Cross-Modal Translation The accompanying position paper is provided as a standalone PDF with a self-contained cover page containing the title, version, author, publication date, and DOI. ARGIRA is a pipeline that sonifies paintings by mapping visual statistics—including color entropy, edge density, fractal dimension, and a 3×3 spatial partition—to acoustic parameters. RMA (Residual Model Auditor) is a computational auditing framework that evaluates which structural information survives a mapping by testing whether characteristics of the original input can still be recovered from its output. This release (v1.3) extends v1.2 by completing Candidate 4's second analytical layer, adding a bootstrap-based stability analysis, giving Structural Texture its own first-class entry in the Candidate Structural Loss Register, and closing the last target shared simultaneously by three candidates. No changes have been made to the frozen ARGIRA v3.4 pipeline or to the audit framework—this release executes previously specified analyses and consolidates already computed results under their owning candidate. Experimental Results Structural Texture (Candidate 2) Structural Texture is given its own full entry in the Candidate Structural Loss Register for the first time. This is a consolidation, not a new experiment: both of this candidate's owned outputs had already been audited to closure under other candidates' entries and are re-attributed here for traceability. Both outputs are confirmed as deterministic, closed-form transforms with no residual possible by construction. Color Structure (Candidate 3) — tempo_bpm Closed tempo_bpm, the first acoustic output identified as being shared simultaneously by three candidates, now has its controlled experiment executed and closed. A nested per-channel model reaches exact recovery, independently confirmed through direct formula reconstruction across the full corpus. This closes the last remaining target shared by three candidates. The previous shared-target case involving two candidates had already been closed in v1.2. Global Structural Representation (Candidate 4) — Layer 2 Executed The second analytical layer, specified but not executed in v1.2, has now been completed in full. The executed analyses include: Correlation matrix Mutual Information Principal Component Analysis (PCA) Global Random Forest feature importance computed across the complete corpus. A fifth analysis—sensitivity analysis under small input perturbations—is introduced in this release, with a documented exclusion of corpus rows where saturated channels make reconstruction unreliable. The pre-registered confounder table introduced in v1.2 has also been expanded after identifying two omissions in the original pre-registration. Both newly added confounder pairs collapse to near-zero correlation once their shared inputs are controlled and are explicitly documented as pre-registration gaps, not as newly discovered pipeline behavior. Two additional acoustic output pairs were found to correlate despite sharing no pipeline inputs. These correlations are traced to relationships among the visual descriptors themselves within this corpus. The documentation explicitly notes that this is a property of the dataset, not evidence of a general pipeline property. Finally, one remaining confounder pair underwent additional auditing and was resolved as a linear-model misspecification artifact, rather than evidence for an undocumented computational pathway. Following the project's auditing protocol, higher-cost confirmatory analyses were judged unnecessary and are explicitly flagged rather than silently omitted. Global Structural Representation (Candidate 4) — Stability Analysis (New) This release introduces a bootstrap-based stability analysis for Candidate 4. Rather than reporting only point estimates, the stability analysis evaluates the robustness of Layer 2 under repeated resampling. Bootstrap analyses are performed for: the complete correlation matrix; confounder-pair partial correlations; global Random Forest feature importances; PCA components and explained variance. For every reported quantity, the analysis provides: bootstrap mean; standard deviation; 95% confidence interval; coefficient of variation; and, for signed quantities, the proportion of bootstrap samples preserving the observed sign. This analysis reveals a result that had not previously been quantified. Among the confounder pairs whose partial correlations were reported as near zero in Layer 2, only one remains consistently stable under bootstrap resampling. The remaining pairs are statistically indistinguishable from sampling noise at the current resampling rate, one of them particularly so. Importantly, this does not reopen any previously closed Layer 2 result. All original point estimates remain near zero as originally reported. The bootstrap analysis simply adds an explicit measure of statistical stability that point estimates alone cannot provide. A companion interactive standalone HTML visualization is also included, displaying the complete source-verified visual→acoustic dependency network, with edge thickness proportional to measured sensitivity. Scope of the Evidence Following Principle 1 of the accompanying research programme, all reported findings constitute either Layer-1 (source-verified structural mapping) or Level-2 (computational) evidence. These evidence types are intentionally kept separate. A source-verified dependency is a fact about the pipeline implementation. A computational result is empirical evidence generated by an RMA experiment. The existence of a dependency does not, by itself, demonstrate information preservation or information loss. The corresponding experiment must still be performed. Level-3 perceptual validation has not yet been performed for any candidate. The documentation explicitly separates: Verified pipeline facts RMA experimental results Interpretation and remaining open questions A VALIDATION_LOG is included documenting material reviewed during the audit that was found inconsistent with the frozen pipeline implementation. Such material is intentionally kept separate from the Candidate Register so that rejected drafts cannot be mistaken for alternative official results. Current Programme Status Step 1 — Blocked by unavailable historical training data. Step 2 — Active and cumulative. Candidate 4's Layer 2 and Stability Analysis are now complete; Structural Texture has its own first-class Register entry; tempo_bpm, the final target shared simultaneously by three candidates, has been closed. All remaining v1.2 results are carried forward unchanged. Step 3 — Perceptual validation has not yet begun. Step 4 — Pipeline redesign remains intentionally frozen pending completion of the audit programme. Files Included zenodo_package_v1.3.zip — Complete artifact bundle. ARGIRA_RMA_research_programme.pdf — Standalone position paper. README.md — Repository documentation. CHANGELOG.md — Project changelog. VERSION.txt STATUS_SCHEMA.md — Validation-status workflow, Evidence Level definitions, and export-only annotations. dependency_mapping_protocol.md — Formal specification of Step 2.0. candidate_structural_loss_register.md — Per-candidate status and audit results. candidate3_analysis.py — Reproducible Color Structure analysis, including the controlled tempo_bpm experiment. candidate3_color_structure_summary.csv — Per-output results table for Color Structure. candidate4_global_representation.md — Global Dependency Map with Layer 1 and Layer 2 results. candidate4_analysis.py — Reproducible Layer 2 analysis for the Global Structural Representation. candidate4_bootstrap.py — Reproducible bootstrap stability analysis for Candidate 4. candidate4_bootstrap_correlations.csv candidate4_bootstrap_partial.csv candidate4_bootstrap_rf.csv candidate4_bootstrap_pca.csv — Raw bootstrap outputs supporting the stability analysis. candidate4_bootstrap_report.md — Human-readable stability analysis report. argiraDependencyNetwork.html — Interactive standalone visual→acoustic dependency network. VALIDATION_LOG.md — Traceability record for reviewed but excluded material. Bitacora.md — Complete methodological development log. sonify_painting_v3_4.py — Frozen ARGIRA pipeline. rma_core_v3.py — Residual Model Auditor implementation. sonificacion_resultados.csv — Corpus (N = 227 paintings). Supporting materials carried forward from v1.0 / v1.1 / v1.2: spatial_arrangement_template.md step1_report.md zones_extracted.json run_design_a_round2.py run_design_a_round3.py run_design_a_final.py run_design_a2.py contexto_proxima_sesion.md v4_redesign_proposal.md The accompanying position paper is included both as a standalone PDF and as part of the complete artifact bundle. Version Repository version: 3.4-audit-1 Zenodo release: v1.3 (DOI: 10.5281/zenodo.21154398) Supersedes: v1.2 (DOI: 10.5281/zenodo.21144059) Concept DOI (all versions): 10.5281/zenodo.21131137 The computational methodology and audit framework remain unchanged. No changes have been made to the frozen pipeline or to any previously computed results. This release extends the audit by completing Candidate 4's Layer 2, adding a bootstrap-based stability analysis, introducing an interactive dependency-network visualization, consolidating Structural Texture under its own Register entry, and closing the final target shared simultaneously by