OpenReliOps: A Safety-Constrained Architecture for Autonomous Reliability in Open-Weight Model Services

Open-weight language models enable private deployment, local adaptation, and independent audit, but they also transfer responsibility for model artifacts, inference infrastructure, telemetry, security, and generated actions to the deploying organization. This paper presents OpenReliOps, a self-healing control architecture that treats an open-weight model as a replaceable reasoning component inside an evidence-grounded and policy-bounded reliability loop. The architecture integrates service-level-objective forecasting, topology-aware telemetry retrieval, causal root-cause diagnosis, a typed repair domain-specific language, independent policy verification, calibrated escalation, progressive execution, rollback, and post-action semantic and operational validation. A reliability process preference optimization objective is defined to learn from successful, rejected, failed, and rolled-back incident trajectories without rewarding unsafe outcome-only behavior. Formal analysis establishes policy confinement and blast-radius bounds under complete mediation, current state, typed actions, and sound declared policies. The evaluation protocol separates diagnostic correctness, evidence grounding, policy safety, and measured service recovery, and specifies comparisons on RCAEval, OpenRCA, OpenRCA 2.0, AIOpsLab, and controlled model-serving faults. The analytical result is that model substitution alone cannot provide autonomous reliability: the reliability boundary is created by the interfaces among evidence, uncertainty, policy, execution, and verification. The framework is therefore positioned as a falsifiable methodological contribution whose operational gains must be demonstrated through incident-level experiments, ablations, multiple seeds, and complete failure reporting.