Loose-Screw Activation Framework: Learnable Leak as a Controlled Auxiliary Signal Path

This paper introduces the Loose-Screw Activation Framework, an exploratory empirical study that reframes activation function leak from an unwanted artifact to be suppressed into a bounded, learnable, and controlled auxiliary signal path. Rather than allowing parameters to drift unchecked, the framework implements a dual-path architecture that balances a primary "clean" path against a secondary "leak" path via pre-defined initialization splits, specifically evaluating 95/5 and 50/50 structural weight allocations. Through an extensive 30-page experimental sweep using standard MNIST benchmarks across shallow and deep multi-layer convolutional networks, we track layerwise weight evolution, performance shifts, and neural dead-unit resistance. The framework is heavily stress-tested under severe label corruption ranging from 10% up to 90% artificial noise. The empirical findings show that the advantages of the dual-path leak framework are highly regime-dependent. While it does not claim universal superiority over standard baselines in clean settings, it serves as an effective structural insurance policy under extreme label noise (80%). Where traditional ReLU units suffer widespread gradient death and unit collapse, the mixed leak-aware variants preserve stable gradient throughput, maintaining a zero dead-unit rate under chosen activation thresholds. This work characterizes the exact diagnostic behavior and boundary conditions of learnable leak allocation during noisy training regimes. Fully adhering to open science and transparency standards, the manuscript explicitly includes direct links to the originating large language model (LLM) conversation logs and the Google Colab experimental notebooks.