Drift-Localized Order Effects in Temporal Dense Retrieval

Temporal retriever refresh must reconcile examples that add facts with later passages that revise or contradict them, but ordinary curriculum comparisons entangle their order with exposure, replay, and negative mining. We identify the residual sequence effect by materializing a 6.4-million-occurrence multiset containing primary and 20% replay triplets before training and permuting only their occurrence-to-step assignment; every occurrence retains its positive and hard negative, and the loss has no cross-example terms. A frozen six-signal drift card defines response strata, while StrataOrder schedules additive, revision, and conflict phases forward. On Reuters News, versioned API documentation, and Wiki Entities, Contriever-MSMARCO reaches 74.7 Recall@100 in the eight-window high-drift slice versus 70.1 for global random order ($+4.6$, 95% CI $[3.1,6.2]$, $d=0.89$) and 72.4 for random order with explicit time tokens ($+2.3$, $[0.7,3.8]$). Reverse ordering jointly reduces final-state recall and current-over-stale preference and increases forgetting; phase-preserving shuffling lies between Random and full forward order. ColBERTv2 reproduces the high-drift contrast (+4.0 [2.5, 5.5]), while both architectures show only $+0.3$ in the predeclared null slice. The resulting drift–order map turns sequence design from a generic curriculum choice into a measurable property of retrospective refresh: revisions and conflicts identify where fixed-exposure scheduling materially changes retrieval.