Learn-by-Wire Training Control Governance: Bounded Autonomous Training Under Stress for Stability and Efficiency

This paper introduces Learn-by-Wire Guard (LBW-Guard), a bounded autonomous training-control governance layer for language models. It aims to improve training stability and efficiency by observing telemetry and applying bounded control, significantly reducing final perplexity.