Narrative-Dynamical Systems (NDS): A Closed-Loop Architecture for Long-Horizon Autoregressive Decoding via Orthogonal Logit Projection and Dynamic Barriers

Standard autoregressive language models typically generate text in an open-loop fashion, ignoring the accumulation of errors over time. Consequently, despite their local fluency, these systems frequently suffer from long-horizon pathologies such as repetitive loops, diminished lexical diversity, and distributional collapse when relying on truncation-based sampling. To address this, we present Narrative-Dynamical Systems (NDS), a closed-loop decoding architecture that couples a frozen generator with a frozen encoder through a modular pre-sampling logit processor. NDS actively monitors online statistics across three channels—representation drift, token-level redundancy, and distributional concentration—and intervenes only when these signals jointly indicate a transition into a degenerate regime (low-drift/high-redundancy). The control action is injected directly into the logit space as a combination of (i) an orthogonally projected ascent step derived from a quadratic KL trust-region surrogate, and (ii) a sparse dynamic barrier designed to suppress empirically identified attractor token sets. We provide explicit derivations for the KL approximation and projection steps, alongside a closed-form bound demonstrating the exponential attenuation of probability mass assigned to the attractor set.