Disclosure Features over Standard Graph Architectures: A Multi-Seed Benchmark of GNNs and a Matched-Input MLP for Post-Filing Fraud Risk Scoring in Chinese A-Share Firms

Reproducibility companion for the PeerJ Computer Science submission. Releases the feature schema/data dictionary, a 50-row synthetic sample, per-seed result JSON, persisted score arrays, and all training/evaluation/table-generation code needed to regenerate the paper's tables and figures. Raw CSMAR feature matrices and train/validation/test split indices are withheld under licensed-data restrictions.