GRAFT-Rec (Graph-Augmented Fusion with Gradient-Boosted Recommendation) is a hybrid meta-ranking framework that combines heterogeneous collaborative, graph-based, content-derived, and popularity-aware signals for top-K recommendation. In response to the reviewers, the revised framework removes outdated language-model prompting and LightGCN-fusion descriptions, integrates RP3\(\beta\) as an internal graph-based candidate and residual signal, and selects the final architecture exclusively by validation performance. The final MovieLens-1 M configuration uses a NoBiRank meta-feature variant with RP3\(\beta\) residual blending and activity-stratified adaptive mixing. On MovieLens-1 M, GRAFT-Rec achieves P@10 = 0.0108, HR@10 = 0.1080, NDCG@10 = 0.0570, and MRR@10 = 0.0419 on 1,037 held-out users, outperforming RP3\(\beta\) (NDCG@10 = 0.0396) by 43.95% and UserKNN (0.0389) by 46.55%. Paired user-level tests confirm that the NDCG@10 advantage over RP3\(\beta\) and UserKNN is statistically supported by paired-difference confidence intervals excluding zero. A secondary MovieLens-100K experiment provides a conservative generalization check: GRAFT-Rec improves over RP3\(\beta\) and UserKNN, but ItemKNN remains the highest-NDCG method on that smaller dataset. A strict global-time audit further evaluates sensitivity to cross-user temporal leakage and shows GRAFT-Rec as the numerically strongest method under a causal time-prefix split. The revised analyses moderate unsupported claims: cold-start and tail-item improvements are not presented as core contributions, while catalog coverage is retained as a supported exposure benefit.
