RS4 Adaptive Prefilter with Offline Sleep-Like Feedback. A closed-loop architecture for source selection before modal dynamics

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The RS4/Modal-Dynamics architecture separates local input integrity, persistent modal state, reality verification, temporal trust, and authority. The present work adds one narrowly defined layer before RS4_F: a budgeted adaptive prefilter whose source weights are updated only after an offline consolidation cycle using independently verified downstream outcomes. The purpose is not to model biological sleep literally, but to test whether a sleep-like feedback loop can improve the quality of information presented to the already frozen RS4_F module. A preregistered-style synthetic benchmark used 12 input sources (6 reliable, 3 conditionally spoofable, 3 noise sources), a fixed prefilter budget of 6 sources, 100 pilot seeds for learning-rate selection, and 300 disjoint confirmatory seeds per method in two suites. With fixed source roles, full offline feedback achieved reliable-source selection precision of 1.000 versus 0.497 without feedback and 0.509 with shuffled feedback. Mean phase error under coordinated spoofing fell from 0.422 rad without feedback to 0.059 rad with full feedback, and early real-change tracking error fell from 0.641 to 0.293 rad. Paired bootstrap confidence intervals excluded zero for all three effects. A shuffled-feedback control remained near baseline, indicating that the benefit depended on correct feedback assignment rather than generic weight perturbation. These results establish a synthetic engineering proof-of-function for a closed feedback loop from verified offline processing to next-cycle prefiltering. They do not establish biological sleep mechanisms, real-world adversarial robustness, or the correctness of the downstream Reality layer. Keywords: RS4; adaptive prefilter; offline consolidation; sleep-like feedback; source reliability; modal dynamics; synthetic benchmark; AI memory