{"id":"arxiv/2610.19207","url":"https://arxiv.org/abs/2610.19207","tags":["math.ST","cs.AI"],"score":0,"title":"Finite-sample certification of coverage regret for budgeted document selection under adaptive stratified auditing","domain":"math.ST","source":"arxiv","authors":["Nora Serdyukova"],"reviews":0,"subject":"math.ST","summary":"Selecting a budget-constrained set of documentary fragments, for instance the knowledge base of a retrieval-augmented generation system, raises a statistical question: how much expected query coverage is lost relative to the best feasible selection when the query distribution is unknown? We construct finite-sample, distribution-free upper confidence bounds for this coverage regret that remain valid after data-dependent selection. A pilot stage may select the document set, screen competitors, build comparison networks and learn an annotation policy. Then an independent audit acquires…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-13","reactions":{}}