S11-批判性审查

Run a structured review that randomly samples 5-10% of both excluded and retained records, checks verdict consistency across similar papers, and compares actual noise rates against pre-cleaning expectations. The process ends with a graded quality report.

Randomly sample 5-10% of records marked as noise and manually verify whether any are actually relevant. Also check borderline cases that moved from pending to excluded, since overly strict handling there is a common source of false positives.

A quick review runs after each cleaning round: it tallies verdict counts, compares with prior rounds, and spot-checks about 40 records. A deep review runs after all cleaning finishes and applies the full six-dimension framework with a final A-to-D quality grade.

A redo is warranted when the audit finds fundamental problems, rated D-grade, such as contradictory rules applied across batches or large systematic deviation from expected noise rates. B or C grades only need documented caveats or targeted fixes.

The expected-versus-actual dimension depends on predictions from the earlier estimation step, but the other five dimensions work independently. Without predictions, that dimension is skipped and the remaining checks still produce a valid quality grade.