Orthogonal Signals for Goal Integrity and Dynamical Stability in RL: Linking OTIP Interventions with NSI via the Variety≠Stability Motif

This release contains the preprint “Orthogonal Signals for Goal Integrity and Dynamical Stability in RL: Linking OTIP Interventions with NSI via the Variety–Stability Motif” by Karel Hrubec (Independent Researcher). The paper reports empirical results on the OTIP Phase3b2 benchmark, focusing on proxy exploitation and how it manifests as rare but large (“heavy-tailed”) return-gap events. Beyond mean metrics, the study explicitly quantifies tail risk using VaR/CVaR, demonstrating that baseline Q-learning can exhibit occasional extreme proxy-aligned outcomes despite a small median gap. An ontology-based intervention eliminates probing/energy-kill behaviors in L2 proxy episodes, achieves perfect success in the evaluated setting, and collapses proxy–control return-gap tail risk. All runs used in the final analysis were produced from a clean output directory and passed a commit-chain integrity audit (80/80 runs). To connect integrity-oriented measures with dynamical signals, the release also includes an NSI-style stability index computed from per-step traces (entropy-based “variety” and dispersion-based “stability”), illustrating that goal integrity and dynamical stability are not redundant (stability can remain high even when goal success collapses under ablation). Contents PDF of the preprint. Overleaf/LaTeX source (tables and appendix included). Aggregated result tables for Phase3b2 (final clean) and optional parameter sweep outputs (where applicable). Keywords: reinforcement learning safety, proxy exploitation, reward hacking, tail risk, CVaR, robustness evaluation, benchmark, OTIP, NSI, dynamical stability, goal integrity, reproducibility.