Record-held or activity-held? A preregistered pause test and weight-edit test on self-optimizing networks, and a GPU test of near-tie decisions in a continually learning language agent

This preregistration fixes, before any data are collected, the design of two studies. Study A applies the pause test to resistance itself, on the self-optimization Hopfield model (Froese et al. 2023; Weber, Guckelsberger and Froese 2025): resistance that survives a halt of the system's activity is held by a record, while resistance that fades when activity stops depended on the system's ongoing activity. It adds an outside edit of the learned weights and measures whether reset-converge-learn dynamics restore it across learning-rate regimes. Study B reruns two near-tie decisions of a continually learning language agent (Yuan 2026, The Agent Is Its Record) on a GPU under varied batch composition and counts how often hardware nondeterminism flips the choice. Predictions and refutation criteria are stated for each study.