G-XCI: A Conditional Multi-View Audit of Stability and Cross-View Rank Agreement in Tabular Feature Attributions

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Post hoc feature attributions are often interpreted from a single run, although their reproducibility and agreement with other explanatory views can depend on resampling, model family, representation, audit lens, output orientation, and inspection scope. We evaluate G-XCI as a conditional three-readout audit protocol. LAS measures reproducibility of an aggregate ranking derived from held-out Monte-Carlo Shapley attributions; GRS summarizes cross-seed variance of a specified global score profile; SCC measures Kendall tau-b agreement between the aggregate attribution ranking and a chosen global ranking. Across three tabular datasets, three model families, five stratified repeats, four global audit lenses, three inspection scopes, GRA ablations, and representation variants, Adult Income changed SCC sign with K, while audit-lens and representation changes materially altered SCC. Targeted supplementary audits showed that complementing the WDBC binary output reversed the mean GRA-based SCC sign in all nine model-scale combinations while leaving absolute local Shapley rankings unchanged, whereas forcing Standard GRA, absolute Spearman-to-output, and output permutation onto the same 128 held-out cases preserved all 27 original mean-SCC signs. A short lemma shows that robust-z followed by min-max equals direct min-max for nonconstant sequences. In the tested tabular binary-classification settings, G-XCI supports conditional auditing of aggregate-ranking reproducibility and cross-view rank agreement under a declared audit contract; it does not establish universal explanation correctness.