ChronoGate: Temporal Admission Before Ranking in Versioned RAG

Temporal admission and finite relevance ranking expose different evidence: ranking inside a query-conditioned set, ${\mathrm{Top}}_k^s(B_a)$, cannot in general be recovered by filtering the globally ranked set, $B_a\cap {\mathrm{Top}}_k^s(C)$. In versioned corpora, this ordering decides whether target-supporting passages survive competition with topically similar but incompatible versions. ChronoGate operationalizes the boundary as a calibrated choice among no retrieval, point-valid support, bounded historical support, and temporally constrained multihop retrieval; the selected set is applied before dense search, while a separate support-monotone gate handles residual decoding conflict. An intervention protocol evaluates this placement on 115,480 records from WikiRevision-Real, GovPolicy-Open, FinNews-Exec, and a TemporalQuestions-derived corpus while holding the downstream reranker, generator, decoding, and metrics fixed. ChronoGate improves temporal consistency by 6.5–7.2 points over RouterEns+NLI and reduces dataset-macro dense searches from 0.95 to 0.71 per query. On WikiRevision-Real, missing-target-support error falls from 6.2% to 3.2%, incompatible-version adoption falls from 5.9% to 4.1%, and removing temporal windowing costs 8.6 points; the consistency margin widens from 4.3 to 10.2 points across the measured update-rate strata. A time-conditioned dual encoder at the same admission locus narrows the aggregate margin to 1.1–1.2 points, but the difference reaches 2.4 points in both very-high-update and high-conflict slices; both official TimeQA splits and a blinded 1,800-query evidence-span audit preserve the ordering. These results identify pre-ranking set selection as the most effective tested temporal-control locus when superseding versions compete for limited ranking capacity.