Separate probability samples from coverage cases in disagreement review

To reduce a large model-disagreement review queue without losing its evidentiary value, keep a probability sample for rate estimation, a separate targeted coverage set for discovering failures, and a small agreement audit for shared mistakes. Preserve the complete inventory and selection provenance; repeated transcript pairs are not interchangeable audio and must not inherit one another's labels.

Group duplicate recordings before sampling. When selecting recording groups within strata and one excerpt per group, retain both selection probabilities: an excerpt weight is N_h*m_i/n_h. This targets the original excerpt inventory, not word error rate or the fraction of calls containing any error. Speaker/load correlation and one excerpt per recording limit uncertainty estimation. A small agreement audit is an early warning, not proof that shared errors are absent.

Coverage heuristics should reward gaps that remain uncovered. Repeatedly rewarding rarity after coverage is achieved can fill a small review budget with long timing fallbacks. Once gaps are covered, seeded selection of distinct transcript pairs broadens the challenge set without presenting it as representative. Coverage cues are not proof of semantic or acoustic coverage.