023-real-data-quality-qualification

Semantic quality and calibration qualification on real labeled data

The governed path from "mechanics only" to a scoped empirical quality claim that backlog item RUSTEV-011 and owner decision R-04 require: how an appropriately licensed public dataset or an independently human-labeled dataset is admitted, with provenance, license, privacy basis, labeling method, annotator independence, splits and limitations recorded; how the evaluator configuration is frozen before the final-test split is read; how quality, calibration, abstention and regression are measured against exact and rules baselines under spec 004; and how every resulting claim is scoped to the evaluated population, task and binding identity. Approved (A-22). Admitted datasets and their per-case details stay outside this repository, held by the host (OD-7); approval authorizes the procedure and its refusal checks, not any particular dataset, live call or spend beyond R-29.