Dependence-Aware Multi-Head Activation Monitoring for Distribution Shift and OOD Reliability

Abstract

Modern neural systems frequently fail under deployment shift because confidence-only diagnostics underrepresent hidden changes in internal activations, and static benchmark metrics do not directly control sequential alert burden. We study a reusable, architecture-agnostic activation-monitoring framework that combines three detector families at each layer: confidence-energy scores, distance-based geometry scores, and streaming discrepancy scores. The method introduces dependence-aware calibration, where global alert thresholds are inflated by online variance and autocorrelation estimates, and a boundary-aware ranking analysis that clarifies when energy-based and discrepancy-based detector orderings agree or diverge. Across static OOD and streaming stress evaluations, fused monitoring improves near-OOD AUROC by 0.013 and far-OOD AUROC by 0.020 over the strongest single-head energy comparator while preserving matched-recall false-alert parity in this evaluation. Time-to-detection improves from 10.87 to 5.40 windows relative to the energy baseline, and boundary-aware selection reduces policy regret by 21.0% on average over five seeds. Symbolic validation verifies 11 of 12 theorem obligations; the kernel-regime inversion result is supported for γ ∈ {1, 2, 4} and fails at γ = 0.5, which explicitly bounds generality. These results support a conditional conclusion: multi-head, dependence-aware activation monitoring improves detection utility and operational interpretability, while full generality still depends on unresolved low-bandwidth and real-data replay regimes.