Benchmark and Item-Level Results for "Can Local Large Language Models Screen and Revise Questionnaire Items? Evidence from a 200-Item Defect Benchmark"

Item-level outputs behind the manuscript "Can Local Large Language Models Screen and Revise Questionnaire Items? Evidence from a 200-Item Defect Benchmark". The archive contains the 200-item gold benchmark (40 clean controls, 160 flawed items under a 16-label defect taxonomy), the labels predicted and revisions produced by a baseline checker-reviser workflow and an orchestrated, role-separated workflow, the run metrics and configurations, and a detection-matched revision analysis with its script. The runs cover Qwen3.5-9B under three prompt conditions and Mistral-7B-Instruct-v0.3 and Gemma-2-9B-it under the first prompt condition. See README.md in the archive for the file layout and the limitations.