STABLE-RAN: Counterfactual Drift Attribution and Stability-Aware Agentic Deep Learning for Verified Model Switching in Open Ran

Abstract

Open Radio Access Network (O-RAN) controllers increasingly depend on deep models whose validity can decay as traffic, mobility, interference, and compute conditions change. Existing drift-handling work detects degradation and retrains models, while emerging agentic architectures propose lifecycle orchestration and rollback. A remaining management problem is deciding which operating cause produced the drift and whether a candidate model should be switched without causing lifecycle oscillation. This paper proposes STABLE-RAN, a verified model-switching framework that combines uncertainty-augmented Page-Hinkley monitoring, counterfactual feature-group attribution, regime-specialist multilayer perceptron ensembles, improvement-gated selection, a minimum dwell interval, pre deployment twin assessment, and versioned rollback. Evaluation uses a controlled normalized O-RAN service-management simulator with five sequential regimes, five random seeds, 2,100 intervals per seed, five lifecycle baselines, and 52,500 method-interval observations. Against a static deep model, STABLE-RAN reduces action mean absolute error from 0.0964 to 0.0351 and the SLA-violation rate from 0.4974 to 0.2397. Against an unconstrained agentic switcher, it reduces mean switches from 5.2 to 4.2 and oscillations from 1.6 to 0.4, but incurs higher error and a 4.76-percentage-point SLA penalty. Attribution accuracy is 0.773. In 800 independent candidate-corruption trials, all unsafe candidates are detected, benign false rollback is zero, and rollback lowers corrupted-candidate SLA violations from 0.330 to 0.245. The results establish a measurable stability–adaptation trade-off in controlled simulation; they are not carrier, hardware, or 3GPP-conformance measurements.