Project GlassBox: Structure Over Scale in Neural Reasoning — Breaking the Black Box Through Architectural Transparency and Antifragile Adaptation

Project GlassBox is a systematic 59-phase experimental campaign demonstrating that small, structurally constrained neural architectures can simultaneously achieve superior task performance and unprecedented interpretability compared to large unconstrained models. Using ARC-AGI as a benchmark for abstract visual reasoning, a 77K-parameter Graph Neural Network with Pointer attention (the "GlassBox Agent") outperforms a 1.45M-parameter Transformer baseline (56.8% vs 43.9% full match accuracy). Through test-time gradient adaptation with geometric data augmentation, accuracy reaches 87.4%, breaking through a previously observed 85% performance ceiling. What's new in v2: Antifragile Intelligence (Phases 34–50): Gradient-based ablation achieves "super-compensation" — deliberately destroying low-importance weights improves post-adaptation performance. 10 optimization strategies tested; simplicity consistently wins. Neural Metabolism (Phases 51–55): Combined ablation + neurogenesis reaches 90.8% peak accuracy (single seed) via fine-grained grid search over ablation/neurogenesis rates. Statistical Reckoning (Phases 58–59): Rigorous multi-seed validation reveals the 90.8% peak is not statistically significant (p=0.849). The true benefit of ablation is variance reduction: 15% ablation reduces performance std from 3.0% to 0.7% (4.3× reduction), making the system dramatically more reliable. Key Results: Structure > Scale: 77K structured parameters outperform 1.45M unstructured parameters (19× smaller, higher accuracy) Hydra Self-Repair: First quantitative characterization of neural self-repair — after destroying 50% of model neurons, few-shot adaptation recovers 95.8% of original performance 82.8% Attribution: Full causal path tracing for 82.8% of predictions, exceeding by 3.3× the 25% attribution coverage reported for large language models Ablation as Variance Regularizer: Gradient-based ablation at 12–15% reduces seed-dependent variance by 4–5×, transforming ablation from a performance booster into a reliability mechanism Source code: https://github.com/hafufu-stack/glassbox Acknowledgments This research was conducted entirely independently, without institutional affiliation or corporate funding. The author currently faces financial constraints that make it increasingly difficult to maintain subscriptions to AI services essential for this line of research. To sustain and improve the quality of future work, the author is actively seeking community sponsorship. Details are available at https://github.com/sponsors/hafufu-stack.