Id: arxiv/2611.22096 Url: https://arxiv.org/abs/2611.22096 Tags: stat.ME, math.ST, stat.ML Score: 0 Title: COSI: A Framework for Robustly Valid Inference Domain: stat.ME Source: arxiv Authors: Arun K. Kuchibhotla, Sivaraman Balakrishnan, Larry Wasserman Reviews: 0 Subject: stat.ME Category: research Comments: 0 Disputed: false Division: stat Mentions: 0 Discussed: 0 Published: 2026-11-17 Page: https://delphi.science/r/arxiv/2611.22096 We introduce COSI (COnfidence from Scale Invariance), a framework for constructing confidence sets for statistical functionals. COSI builds on the HulC (Hull-based Confidence), which uses median unbiasedness to construct confidence sets from batch estimates. COSI uses scale-invariance properties shared by many common estimators. We divide the data into batches, compute an estimate in each batch, and retain each candidate parameter value for which the batch estimates pass a test of scale invariance. The resulting confidence sets are asymptotically valid under weaker assumptions than classical…