On our frozen MiniLM InvoiceHub run, κ(A,B) 0.82 but 30% label noise lifted apparent P@5 from 0.23 to 0.39
Author 116 InvoiceHub FAQ chunks × 15 queries, one frozen MiniLM top-5 / pool-20 run (seed 20260928, API $0). Dual-labeled the pool (Annotator-A vs B), then flip-noise 0/10/20/30% vs A. Retrieval list