ArtCreaS a Decision-First Assessment Architecture for Art Education

Fluent AI-generated critique does not by itself establish educational validity. ArtCreaS links construct-overlap diagnostics, missing-evidence handling, and uncertainty reporting to a proposed teacher-controlled feedback workflow. The evaluated assessment core uses ExtraTrees and ridge regression; no generative model or classroom feedback intervention was evaluated. The primary benchmark comprises 3819 creativity-labeled APDDv2 records, including 1877 with comments. Across five stratified 80/20 splits, Reliability-Gated, Craft-adjusted Multimodal Fusion (RGC-MF) achieved MAE = 0.306 ± 0.007, Pearson r = 0.954, and R2 = 0.910 on the 1–10 score scale. The 0.002 MAE difference from structured ExtraTrees does not establish a meaningful or statistically supported improvement. At 50% independent rubric dropout, five-split, three-mask evaluation yielded MAE = 0.506 ± 0.009 for RGC-MF and 0.593 ± 0.011 for early fusion. A pooled out-of-fold residual interval attained mean coverage of 88.1% overall and 83.8% for sketches at a nominal 90% target. The stress tests concern controlled random evidence loss, while the fixed-width interval is a descriptive heuristic rather than standard split conformal prediction. Public AuDrA outputs provide a separate visual-scoring anchor (r = 0.802), rather than external validation of RGC-MF. The contribution is an assessment architecture supported by repeated prediction, robustness, and subgroup analyses. Generative wording, teacher release, student contestation, and learning effects remain proposed validation targets.