YongleAI-SciData: Multi-engine candidate annotations and metadata for Yongle Palace Taoist mural image patches

This dataset contains the public image, annotation, and metadata layers for a patch-level candidate-annotation resource derived from five panoramic photographs of the Chaoyuan Tu murals in the Sanqing Hall of Yongle Palace, Shanxi, China. The release includes 7,455 derived 512 x 512 pixel patch PNG files, patch manifests, source-level split metadata, identifier crosswalks, STEP3 per-engine candidate annotations from four vision-language model lines, rule-based consensus and ambiguity metadata, auxiliary STEP4 damage-triage tables, controlled vocabularies, prompt and engine-configuration snapshots, agreement statistics, validation documentation, checksums, and citation metadata.The release also includes a human-review stimulus archive containing 200 red-box 3 × 3 context views, 200 target patch crops, four contact sheets, and a red-box geometry manifest used for the two-expert review-control layer. The annotations are machine-generated candidate labels rather than adjudicated reference labels or benchmark-grade ground truth. Tie indicators, consensus-status fields, and review-priority metadata are provided to support uncertainty-aware reuse and later human curation. The original full-resolution source panoramas are not redistributed in this package because they remain under separate institutional rights governance. The public CC BY 4.0 license applies to the released derived patch PNG files and annotation-and-metadata tables bundled in this Zenodo record.The human-review records are provided as a review-control and validation-support layer, not as a complete adjudicated ground-truth label set for all 7,455 patches.