ONLINE RESOURCE 1 - SCFCE-R EXPANDED REPRODUCIBILITY PACKAGE

Target journal: Social Network Analysis and Mining
Manuscript: SCFCE-R: Validation-Tuned Risk-Gated Graph Attention for Budgeted Misinformation Diffusion Control

This archive contains all original experiment artifacts plus the expanded journal benchmark suite.

FINAL FROZEN POLICY
If supplied content risk r_c < 0.55: abstain.
Otherwise rank nodes by:
  0.8 * rank(GAT susceptibility) + 0.1 * rank(betweenness) + 0.1 * rank(degree).

PRIMARY CONFIRMATION
- BA, WS, ER, SBM
- 150 nodes
- 30 replications per topology
- 120 paired graph realizations per budget
- budgets 2%, 5%, 10%

EXPANDED JOURNAL BENCHMARKS
- Prediction: AUROC, AUPRC, F1, precision, recall, MCC, balanced accuracy, Brier, ECE, log loss, Precision/Recall@5% and @10%.
- Structural baselines: random, degree, PageRank, k-core, betweenness, community bridge, eigenvector, Collective Influence, VoteRank.
- OOD topologies: LFR, Holme-Kim/power-law-cluster, random geometric.
- Diffusion transfer: content-conditioned IC, weighted Linear Threshold, uniform IC p=0.08 and p=0.12.
- Intervention metrics: final reach, spread AUC, peak incidence, containment time, harm reduction, harm reduction per budget, benign utility.
- Statistics: paired Wilcoxon, Holm correction, bootstrap 95% CI, paired rank-biserial effects.
- Community burden: max-min intervention-rate gap, Gini, Jain fairness index, normalized entropy.
- Scale stress test: effectiveness at 300/600/1000 nodes and exploratory multi-scale GAT training.

IMPORTANT CLAIM BOUNDARIES
1. The main evidence is controlled simulation. Empirical social-network structures use simulated diffusion.
2. No observed Twitter15/Twitter16/PHEME misinformation cascade has been evaluated.
3. The sparse uniform-IC p=0.08, 5% budget comparison is not significant after Holm correction.
4. Community-allocation fairness is worse for SCFCE-R than VoteRank/betweenness in the SBM diagnostic.
5. The single-scale frozen GAT loses its intervention advantage in several 600-1000 node stress tests.

Human authors must independently verify the code, references, numerical outputs, interpretations, author metadata, funding, conflicts, and final manuscript before submission.
