Annotation rather than conclusion: data and analysis code for an AI-supported design studio screening study

Dataset and analysis code for a three-arm study of how much an AI assistant is permitted to conclude when it supports teacher screening in an industrial design studio. 240 studio proposals were allocated at random, 40 per brief per arm, to a standard workflow, a conclusion-generating AI that returned a recommendation, or an annotation-only AI that annotated six criterion dimensions without issuing any verdict, score, ranking or recommendation. Each proposal was judged once by each of three teachers, giving 720 judgement units, 240 per condition. The same 240 proposals were independently blind-reviewed by a three-member expert panel, which supplied the reference standard. Supervisor time was logged across the 15-day cycle, seven days of baseline and eight days with the annotation-only system in the workflow. The deposit contains the four collected data tables (proposals, teacher judgements, expert blind reviews, supervisor time log), an English data dictionary, and three Python scripts that reproduce every statistic reported in the article: descriptive statistics, cluster-bootstrap confidence intervals, generalised estimating equation models, standardised effect sizes, equivalence tests and inter-rater reliability. Students appear only as sequential codes S001-S240, teachers as T1-T3 and experts as E1-E3. Free-text proposal titles are withheld because they describe student work in progress. The Teaching Affairs Office and Research Affairs Office of Sichuan University of Science & Engineering classified the study as non-intrusive, record-based and low-risk research and issued a written ethics exemption; no approval number applies.