Self-Correction Emerges from Recurrent Structure in a Minimal 35-Neuron Network: Ablation Evidence

Published version (peer-reviewed, revised): "The Fake Mirror Effect: Foreign Feedback Disrupts Self-Correction in Minimal Recurrent Networks," Transactions on Machine Learning Research (2026), https://openreview.net/forum?id=cwENvGCLRv. The published version supersedes this preprint; the title, claim scope, and several reported values were revised during review. We construct a minimal recurrent neural network of 35 neurons and demonstrate that self-correction — the ability to revise predictions using output feedback — emerges under appropriate learning conditions and is entirely dependent on the recurrent self-referential loop. Ablating the recurrent pathway eliminates self-correction while preserving baseline accuracy. Replacing veridical self-feedback with shuffled or clone-model feedback degrades performance below the no-feedback baseline, establishing that self-correction depends on the model's own output trajectory, not just any reasonable signal. A parameter-matched feedforward control rules out capacity effects. Hyperparameter sweep across 80 configurations (800 experiments) confirms robustness. Implemented in pure NumPy (no deep learning frameworks) for full transparency.