Vulnerability and Controllability of Origin-Discrimination Features (AI / Painting / Photograph) in ARGIRA

Summary This repository brings together a continuous line of research investigating whether specific ARGIRA features can distinguish between images generated by artificial intelligence, real paintings, and real photographs. The results show that the signals studied do not constitute reliable evidence of image origin under the experimental conditions analyzed: fractal_D does not significantly discriminate between AI-generated images and paintings when using the actual production implementation and a balanced sample. luminance_contrast + sat_mean shows statistical separation between groups, but is vulnerable to common modifications of brightness, contrast, and saturation. The production z_combined score can be shifted through controlled photographic edits without the structural characteristics of the image shifting in the same way. The stress tests show within-class score variability comparable to the separation observed between classes. Scientific Scope This repository documents the vulnerability, limitations, and controllability of specific signals used or considered by ARGIRA. It does not propose a new AI-image detector, does not evaluate machine-learning models, and does not allow the origin of an individual image to be determined. The results should be interpreted within the corpus, generators, and experimental conditions described in the included materials. Experimental Chain The repository brings together nine documented studies or experimental phases with verifiable artifacts: Exploratory test of fractal_D and candidate features, culminating in Round 5 with 20 AI-generated images versus 20 paintings, using a faithful port of the actual ARGIRA v3.5 production implementation. Separation using luminance_contrast + sat_mean. Validation using development and validation splits. Cross-check using images generated with Gemini. Experimental control of illumination, brightness, contrast, and saturation. Audit of luminance_mean. Vulnerability control using real paintings. Systematic stress test with 13 perturbation families applied to three base images. Expanded stress test using nine base images, including AI-generated images, paintings, and photographs. The reproducible artifacts from these phases are included in the repository. Main Result The accumulated evidence indicates that global features based primarily on luminance and saturation can produce statistical differences between image sets without those differences constituting a robust property of image origin. In particular, their sensitivity to common photographic transformations prevents them from being interpreted as reliable indicators of provenance under the conditions studied. This result complements and further qualifies the limitations already documented in ARGIRA v1.4.4. What This Study Does Not Demonstrate This repository does not, by itself: establish the origin of any individual image with certainty; constitute a complete AI-image detection system; evaluate every existing AI image generator or image-generation pipeline; evaluate machine-learning architectures as alternative approaches; establish that the observed behavior necessarily generalizes to all possible datasets, generators, or imaging conditions. Importantly, these limitations do not invalidate the findings of this study. Rather, they define their scope: the experiments provide evidence that the specific ARGIRA signals examined here can be affected by common image transformations and therefore should not be treated as standalone evidence of image origin under the tested conditions. The natural-match and edit-based feature-matching phases (10C/10A) are not included as reproducible results in this repository because no corresponding verifiable artifacts were located. Relationship to ARGIRA v1.4.4 This repository constitutes a scientific companion study to ARGIRA v1.4.4, previously published, and provides additional experimental evidence concerning a limitation that was already documented in that version. ARGIRA v1.4.4 DOI: https://doi.org/10.5281/zenodo.21924218 This study is an independent scientific repository and is not part of the ARGIRA v1.4.4 release itself. Intended Application in ARGIRA v1.4.5 The results of this repository are intended to inform the evolution of ARGIRA toward version 1.4.5. Publication of this study and implementation in the HTML are separate decisions: first, the scientific evidence is established and published in a traceable form; subsequently, it will be determined which specific conclusions should be incorporated into the interface and logic of ARGIRA v1.4.5. Therefore, this repository does not itself constitute the implementation of ARGIRA v1.4.5. Its results will be used as external evidence to determine and justify the changes incorporated into that release. The future ARGIRA v1.4.5 release should cite this repository through its DOI as part of the scientific traceability of modifications derived from these results. Reproducible Material The repository includes: code used in the analyses; ARGIRA v3.5 source code used to verify the port; experimental reports; raw data and aggregated results; reproducibility packages for the stress tests; the image corpus used in the studies; provenance and contextual documentation; MD5 checksums for verifying the integrity of the distributed files. Repository Structure The complete reproducibility material is distributed as a single ZIP archive. Because Zenodo displays the ZIP as a single file rather than exposing its internal directory structure as individually browsable files, the archive preserves the complete verified structure of the research materials. The archive contains: scripts/ — production-code ports, source code, and analysis scripts; informes/ — experimental reports documenting the main research phases; datos_08_stress_test/ — raw and aggregated data from the systematic stress tests; datos_soporte/ — supporting feature data and audit results; experimento_09/ — reproducibility material for the expanded stress test; corpus_imagenes/ — the image corpora used in the studies; contexto/ — contextual documentation related to the research line; README.md — scientific documentation and citation information; CITATION.cff — machine-readable citation metadata; MANIFEST.md5 — MD5 integrity manifest for the distributed files. The ZIP archive therefore constitutes the complete reproducibility package rather than merely an auxiliary attachment. Repository Identification Title: Vulnerability and Controllability of Origin-Discrimination Features (AI / Painting / Photograph) in ARGIRA Author: Jose Ranero García Year: 2026 DOI: 10.5281/zenodo.21945633 License: CC BY-NC-SA 4.0 Relationship: Scientific supplement to ARGIRA v1.4.4 and external evidence for the future implementation of ARGIRA v1.4.5.