DepthLens: Measuring How Evidence Assembly Changes Useful Retrieval Depth

A RAG score observed at one retrieval depth does not identify how long additional evidence remains useful: the assembly operator can change the entire depth–utility response. We define useful-depth shift as the paired displacement of the fitted low-return transition under an assembly replacement, and instantiate it in DepthLens by crossing seven depths with assembly operators on a fixed retriever–reranker–generator stack and resampling matched queries through the complete estimator. On HotpotQA, 2WikiMultiHopQA, MuSiQue, and FEVER, a Manual Ledger that carries entities, supporting evidence, and unresolved hops moves this transition from 7.4–16.2 passages to 3.4–5.2, a 2.2–3.1\(\times\) shift; Qwen2.5-72B-Instruct reproduces 2.1–2.9\(\times\). In the operator ladder, layout-matched and random-schema controls retain CoT-like transitions, whereas a frozen automatic scaffold recovers most of the movement; the Manual-over-Auto residual rises across pooled hop strata, and the Manual-over-CoT gap widens with distractor rate. The measured displacement aligns with held-out action quality: with Manual Ledger fixed, DepthLens reaches 88.9–93.1% near-SLO pair accuracy, stays within 0.5–0.9 primary-score points (EM; FEVER label accuracy) of the feasible oracle, and improves on the matched Learned-\(k_r\) comparator by 0.9–2.9 points. Joint depth–assembly selection then gains 3.2–12.2 points over matched depth-only adaptation at 1.48–1.68 s P95, showing that retrieval evaluation must cross depth with evidence assembly to expose both scientific attribution and efficient operating points.