One Signal Theory - A Conceptual Architecture for Context Management in Large Language Models via Synthetic Affective Signal Modulation of Attention

Abstract

Context window capacity remains a central bottleneck for large language models (LLMs): enlarging the window does not guarantee that a model attends to the right information, and empirical work has shown systematic degradation when relevant content sits in the middle of a long context. Biological cognition faces an analogous problem—an effectively unbounded stream of multimodal sensory input arriving at a working memory of limited capacity—and appears to manage it, in part, through affective and endocrine signals that act as a salience filter, bias- ing which stimuli reach conscious attention. This paper introduces One Signal Theory (OST), a conceptual software architecture that draws a loose analogy to this biological mechanism: a background “subconscious” sub-agent produces low-dimensional synthetic affective indices (here named synthetic adrenaline and synthetic dopamine), converted into multiplicative modifiers— an inverse-temperature term and a pruning threshold—applied to the attention-weight distri- bution of a main “working memory” context. We present the architecture, its mathematical formulation, and probe its behavior on a controlled synthetic salience-retrieval task. Two re- sults emerge. First, we prove, and confirm empirically without exception, that OST’s gating mechanism is rank-invariant: because it is built entirely from order-preserving transforms (tem- perature scaling, thresholding, renormalization), it cannot change which single token receives the most attention, and therefore cannot on its own correct the kind of attention misdirection documented in the long-context literature. Second, we show that OST’s originally specified threshold discards the correct signal increasingly often as context length grows (27% to 69% of trials across n=10 to n=1000, for a weak signal); a revised, scale-invariant threshold based on the z-score of raw attention logits removes this dependence only once the driving affective signal is decoupled from raw context length, after which the false-discard rate is flat (∼15–18%) re- gardless of n. We conclude that OST’s original framing as a fix for attention misdirection is not supported by this analysis, but a narrower reformulation—as a context-length-invariant denois- ing layer that sharpens an already correctly ranked signal without disproportionately discarding it—remains plausible, and we specify the calibration under which that holds.