Correcting Citation Inflation via Bayesian Knowledge Integration: A Fully Computational Validation Framework

Traditional bibliometric indicators such as the h-index, g-index, and Category Normalized Citation Impact (CNCI) conflate genuine scientific quality with citation volume, leaving them vulnerable to citation inflation, self-citation, and coordinated manipulation ("citation cartels"). This paper introduces the Bayesian Knowledge Integration Index (BKII), a hierarchical Bayesian metric that models paper-level citation counts with a Negative Binomial likelihood, augmented by Rao Stirling interdisciplinary diversity and functional-citation weighting, and that couples the resulting posterior estimates with an explicit residual-based anomaly-detection layer for identifying manipulative citation behavior. Because independently verified, individual-level real-world bibliometric ground truth (true scientific quality) is fundamentally unobservable, we evaluate BKII through a fully computational validation framework: a synthetic citation ecosystem (N equals 150 simulated researchers, 1,160 papers) is generated from a known latent-quality parameter, a known citation-manipulation mechanism, and known functional-citation and interdisciplinary-diversity covariates. BKII is fitted via a from-scratch, four-chain Metropolis-within-Gibbs hierarchical MCMC sampler implemented in NumPy and SciPy and evaluated against h-index, g-index, i10-index, an m-index proxy, and a CNCI-style field-normalization proxy in recovering the known ground truth. BKII achieves the strongest linear association with true latent quality among all tested metrics (Pearson r equals 0.365 versus 0.081 to 0.342 for competitors), while top-quintile discrimination (AUC equals 0.584) is comparable to, rather than dramatically better than, citation-count-based alternatives (AUC equals 0.555 to 0.584). The framework's principal empirically demonstrated advantage is not the raw index itself but its auxiliary anomaly-detection module, which identifies synthetic citation-cartel membership with AUC equals 0.885. A Saltelli-based global sensitivity analysis shows that population heterogeneity (sigma sub r) dominates the variance of aggregate BKII estimates (first-order Sobol index approximately 0.72), while the Negative Binomial dispersion and functional and diversity weights contribute comparatively little. A separate 450-paper synthetic citation-network sub-study shows that the exact Disruption Index formula recovers a known disruptive and consolidating ground truth well (AUC equals 0.939), whereas a simplified truncated SVD embedding proxy for the Enhanced Disruption Measure does not (AUC equals 0.400), a negative result reported explicitly rather than concealed. All code, synthetic data-generation procedures, random seeds, and numerical results reported in this paper are fully reproducible from the self-contained Python appendix; no claims are made about performance on real bibliometric databases, and this limitation is stated explicitly throughout. No external funding was received.