NetOps-SLM: A Two-Tier Co-Pilot Architecture for Sub-Millisecond Autonomous Network Remediation and Explainable RCA

Modern telecommunication and cloud data center networks demand sub-millisecond Mean Time to Remediation (MTTR) with deterministic guarantees against configuration hallucinations. While generative foundation models possess significant semantic reasoning, zero-shot models suffer from a fundamental "Sysadmin Reflex" — a statistical pretraining bias that attempts to restart software daemons during physical link failures and healthy network states, inducing a 66.7% false-positive actuation rate. To bridge this capability-assurance gap, we present NetOps-SLM, a Two-Tier Co-Pilot Architecture evaluated on authentic Containerlab FRRouting IP/BGP infrastructure. NetOps-SLM decouples deterministic actuation from asynchronous explanatory reasoning: (1) a Tier-1 Fast-Path Actuator that executes bounded, zero-hallucination repairs through a sandboxed TypedActionGate in 0.068 ms (a 164,000x speedup over generative foundation models), and (2) a Tier-2 Explanatory Copilot powered by a 4-bit QLoRA fine-tuned 8B foundation model (Qwen2.5-Coder-7B-Instruct) that distills raw streaming telemetry into compact invariant representations to generate human-grade Root Cause Analysis (RCA) post-mortems for Network Operations Centers (NOCs). Evaluated across 73 live Containerlab failure episodes across five fault families, NetOps-SLM achieves 100.0% schema contract compliance, eliminates out-of-scope subnet mutations, and reduces false-positive actuations to 0.0%. Finally, we present an architectural blueprint for scaling this framework to a 70B Mixture-of-Experts (MoE) model via sparse upcycling for multi-vendor 6G autonomous operations.