KARZOUN-X: A Resource-Aware, Safety-Gated Local LLM and RAG Architecture for Autonomous Spacecraft Fault Diagnosis Under Communication Delay

KARZOUN-X studies a local-first spacecraft fault-diagnosis and low-risk decision-support architecture combining telemetry anomaly detection, local retrieval-augmented generation, a locally deployed language model, deterministic action gating, communication-delay analysis, hard-stress evaluation, and auditable resource measurement. The study separates real SMAP/MSL anomaly-detection evaluation from synthetic diagnosis/action experiments. On separable synthetic cases, retrieved evidence strongly improved expected next-action selection, while a precommitted hard-stress suite exposed substantial failures of epistemic abstention under ambiguous, conflicting, or missing evidence. The work is a research prototype and is not flight-qualified or certified for autonomous spacecraft control.