RevisionScope: Intended-Time Grounding for Versioned Security Advisories

A security citation can entail every generated claim and still prescribe the wrong mitigation: the cited advisory is authentic, but superseded at the time for which the answer must hold. A revision-grounded evaluation makes this failure measurable by pairing each question with an authoritative intended time and an answer-changing revision boundary, applying four interventions at the query, metadata, update, and cache interfaces, and scoring the cited revision directly with judge-free temporal citation validity (TCV). RevisionScope turns the same evaluation object into a scope-carrying inference invariant: one resolved scope, corpus epoch, stable revision chain, applicability state, and cross-revision conflict record govern retrieval and generation. On 2,144 public SecAdvisory and CVEFixes-DB questions, plain RAG attains 91.6 [0.9]% content support while reaching only 27.6% TCV and 73.8% attack success rate (ASR). RevisionScope raises TCV to 92.3% and lowers ASR to 9.4%, contracting the support–scope gap from 64.0 to 3.8 points; the strongest measured source-reliability comparator reaches 85.4 [1.5]% TCV and 14.8 [1.9]% ASR with higher latency and abstention. The operating point also preserves 95.5 [0.8]% temporal answer accuracy with 1.6% abstention on ordinary queries and transfers without threshold tuning to 640 Kubernetes questions at 93.1 [1.0]% TCV and 7.5 [1.2]% ASR. Across 4,288 questions from four versioned corpora, scope-carrying inference raises TCV from 28.4% to 93.2% and lowers ASR from 72.8% to 8.6% at 18.8% P50 latency overhead. Intended-time grounding therefore changes both system selection and the safety meaning of a supported citation.