Structured Radial Residual Geometry in Galaxy Rotation Profiles: Robustness, Hard-Split Generalization, and Falsification Framework (v12)

This archive contains Version 12 of a structured machine learning analysis of radial residual behavior in galaxy rotation curve datasets. The study investigates whether coherent geometric structure exists in rotation curve residuals independent of galaxy identity and explicit acceleration parameterizations. Unlike predictive-only approaches, this work emphasizes adversarial falsification, structural destruction testing, and leakage controls to evaluate whether learned signal reflects genuine radial structure rather than memorization artifacts. All experiments are performed using galaxy-level group-aware cross-validation. This release includes: Cleaned radial dataset with derived residual targets Robustness matrix across feature families Hard-split generalization results Identity leakage controls Structural destruction tests Noise and delay perturbation cycles Final structural attack summaries 3D visualization suite (surface + PCA + UMAP) Reproducible Python scripts 🧪 Scientific Objective The central question examined: Does structured predictive signal exist within radial photometric and kinematic geometry that survives adversarial falsification? Three residual targets were modeled: Absolute Δ residual (absdeltaᵣesidₗog10) Linear residual (residₗinₗog10) MOND residual (residₘondₗog10) Each was subjected to identical robustness and falsification cycles. 🔍 Validation & Falsification Framework To prevent shortcut learning and dataset leakage, the following controls were implemented: 1. Galaxy-Held-Out Cross-Validation GroupKFold and Leave-One-Galaxy-Out splits were used to ensure no radial points from a test galaxy were present in training. 2. Identity Leakage Tests Galaxy-ID-only baseline (R² ≈ 0 or negative) Baseline + Galaxy-ID augmentation (no performance gain) These demonstrate that performance is not driven by per-galaxy memorization. 3. Structural Destruction Tests Within-galaxy feature permutation (radial disordering) produces substantial degradation in predictive performance. This indicates that signal depends on internal radial coherence. 4. Target Randomization Global shuffle of target variable results in full collapse (R² ≈ 0), confirming absence of spurious cross-fold leakage. 5. Noise and Delay Attacks Velocity noise injection (up to 4× observational uncertainty) Radial delay shifts (k = 1–8) Signal degrades progressively under structural corruption while remaining stable under moderate noise. 6. Acceleration Removal Replication Experiments repeated with: Baryonic acceleration terms removed Both baryonic and observed acceleration removed Structural predictive signal persists. 📊 Summary of Observed Behavior Across all targets: Strong generalization under galaxy-held-out validation Collapse under target randomization Degradation under intra-galaxy disorder No gain from identity augmentation Replication across residual definitions These patterns are consistent with coherent radial structural encoding rather than trivial memorization. 📦 Included Materials This archive includes: rarₚointscleanwithdelta. csv Robustness matrices Hard-split summaries Falsification summaries Final structural attack summaries 3D interactive HTML visualizations Reproducible Python scripts All results are reproducible from included code. 🧠 Version History This release represents Version 12 of iterative structural testing. Earlier versions progressively introduced: Leakage controls Acceleration-term ablation Hard OOD splits Structured adversarial attacks 3D manifold analysis v12 represents the finalized defense suite and documentation pass.