# Data Statement

CXR-CurateBench measures curation-policy effects under a frozen grounded chest X-ray evaluation contract. It is not a clinical decision-support dataset and should not be used to infer patient-level performance.

The primary residual contains 318 box-expected rows spanning 186 benchmark source IDs. It is source-ID-disjoint from the visible-rationale supervision pool, MIMIC-heavy, and not verified as patient- or study-disjoint. The supplemental VinDr diagnostic contains 283 rows spanning 159 source IDs and is reused-source, not training-disjoint.

The reviewer-safe upload contains derived metadata, summaries, aggregate endpoint diagnostics, and evaluator/readout files. Raw source images and row-level records remain under upstream access controls.