ORMAS: Neural Architectural Transparency Enables Autonomous Self-Correction

We argue that the opacity of neural networks under training is not an inherent property of deep learning, but a mathemat- ical consequence of global backpropagation, where a single entangled error signal renders mid-training structural failures difficult to detect. We show that structurally isolating each node’s local gradient chain to a bounded 7-operation path pro- vides a parallel local attribution channel that enables training- time structural transparency and autonomous self-correction alongside the standard global gradient. Our architecture, OR- MAS, extends backpropagation with a per-node local loss an- chored through a capacity-constrained shared bottleneck and a health-gated self-correction mechanism that autonomously diagnoses and repairs structural pathologies in real time. Gra- dient conflicts between signals are resolved via PCGrad pro- jection, and all corrections satisfy a mean-centering conser- vation constraint bounded by an Input-to-State Stability (ISS) analysis. The architecture natively emits five layers of struc- tural attribution telemetry as physical properties of the for- ward and backward pass, not post-hoc approximations. Across 383 controlled experiments on four architectures (FC-DAG, CNN, Fat CNN, ResNet-18), ORMAS matches standard base- lines on clean data while limiting accuracy decay under 40% label noise to 2.5 percentage points versus 7.8 for standard training. Under catastrophic mid-training structural collapse, ORMAS recovers to 80.3 ±1.6% where parameter-matched baselines permanently collapse to 10.0%. Under sequential task shift with no replay buffer or modularity constraint, the network achieves 58.8% zero-shot compositional accuracy versus 25% chance—an outcome that arises because gradient conflicts between sequentially learned objectives force spatial separation of competing representations without explicit in- struction. Correction frequency decays from 4.2 to 0.05 per epoch, consistent with the convergence predicted by the ISS stability analysis. These results establish architectural trans- parency as a structural prerequisite for autonomous robust- ness.