RAG needs two gates, not one score

SIRIN treats contextual hallucination as two runtime questions: can the retrieved context answer the query, and does the generated answer stay supported by it? It puts probing, uncertainty estimation, and judge-style verification behind one interface, with response- and span-level inspection. 1

Start with the public SIRIN Space in shadow mode. Feed context-query-answer triples, log answerability and faithfulness decisions, and make serve/retry/abstain the reversible handoff to your current RAG path. In the paper's SimpleMem + LongMemEval run, the two gates lift served accuracy from 62.4% to 79.3% and cut strict unfaithfulness from 20.2% to 9.8%. Those are one-backbone, English research results; calibrate on your own corpus before routing live traffic. 1