The Grounded-but-Wrong Gap in Cybersecurity RAG: Evidence from a Bilingual Benchmark and Multi-Judge Evaluation

Retrieval-augmented generation (RAG) is increasingly deployed in safety-critical cybersecurity question answering, where an answer must be both factually correct and faithful to its sources. Evaluation in this space routinely conflates the two, treating grounding as a proxy for correctness. We show that this conflation is unsafe. We introduce SENTINEL-QA, a bilingual English and Macedonian cybersecurity QA benchmark of 70 hand-authored items over a curated corpus of authoritative source documents, and use it to compare four answer generation pipelines (zero-shot, Classic RAG, DSPy-structured RAG, and GraphRAG) across three generators, judged by a four-model LLM panel on both factual correctness and grounding. The two metrics decouple under graph-based retrieval augmentation: on all three generators our chunk-similarity GraphRAG makes answers look better sourced without making them more correct, and as the retrieval budget grows it can even degrade factual correctness while grounding holds steady. Structured, minimal prompting improves both qualities, at lower latency and cost; a four-cell ablation attributes the factual gain to pruning the instruction, while the declarative schema contributes on grounding. The judge panel mirrors the split, agreeing more on what is well grounded than on what is true, under absolute as well as rank-based agreement criteria. We conclude that grounding alone is an insufficient target for RAG evaluation and that pruning the instruction, rather than expanding retrieval, is what improves factual correctness, without eliminating the grounded-but-wrong regime.