Auxiliary Governed Attention: A Governable, Inference-time Auxiliary Attention Mechanism with Sovereign Boundaries for Frozen Transformers

We introduce Auxiliary Governed Attention (AGA), a post-hoc, inference-time auxiliary attention mechanism designed for frozen pre-trained Transformers. Unlike retrieval-augmented generation or parameter-efficient fine-tuning, AGA operates as a governable side-channel that injects bounded corrective signals into the primary attention output, without modifying model parameters or externalizing knowledge into prompts. AGA is built around the principle of sovereign separation between primary and auxiliary reasoning, enabling strict control over when and how auxiliary information may influence inference. We formalize this control through: (i) a governance framework that distinguishes intrinsic, borrowed, and governed knowledge; (ii) explicit lifecycle semantics for auxiliary learning units (probationary, confirmed, deprecated, quarantined); (iii) a hierarchical rights model regulating activation, contribution, and persistence during inferencedesigned to prevent reasoning drift and ghost knowledge accumulation. In addition, AGA employs primary-entropy-gated intervention, ensuring that auxiliary attention is activated only when the primary model exhibits uncertainty. Experiments on commonsense reasoning (CommonsenseQA, PIQA), long-context question answering (Natural Questions, HotpotQA), and continual knowledge injection tasks show consistent improvements of 2.1-4.3 percentage points with less than 0.6% parameter overhead. Ablation studies demonstrate that governance mechanisms are critical for long-term stability, preventing uncontrolled accumulation of auxiliary knowledge and collapse into implicit retrieval-based behavior.