Mechanistic Transparency of Neural Networks: A Four-Layer Framework for Demystifying the Black Box Authors / Creators

This work investigates whether shallow neural networks exhibit transparent and controllable internal structure through a four-layer mechanistic interpretability framework. Using a 784→16→10 fully connected neural network trained on MNIST, the study evaluates transparency across four dimensions: geometric representations, neuron-level semantics, modular algorithmic circuits, and causal mechanisms. Results demonstrate a 3× improvement in class separability, 87.5% monosemantic neurons, sparse modular circuits averaging 3.1 neurons per class, and successful causal model editing using Rank-One Model Editing (ROME). A targeted intervention restored corrupted class performance from 0% to 97.86% while modifying only 0.28% of model parameters. The findings suggest that shallow neural networks can exhibit interpretable and causally editable internal computations, providing evidence that neural network opacity may partly reflect limitations of analytical tools rather than intrinsic properties of the models themselves. All code, experimental procedures, and reproducibility materials are provided to support independent verification.