<![CDATA[Carbonate fracture characterization typically conditions fracture porosity (Fpor) on lithofacies classification, on the assumption that facies labels capture the geological controls on fracture development. This study evaluates this premise by employing a two-scale statistical decomposition of Fpor variance across 537 samples from 13 wells within a single field in the Malay Basin. The analysis uses FMI-derived Fpor as the reference and the operator's three-class lithofacies scheme (clean limestone, mixed carbonate, tight shaly carbonate) as the facies framework. The lithofacies classification produces a statistically detectable but quantitatively small effect on Fpor (one-way ANOVA F = 3.85, p = 0.022, η² = 0.014; Kruskal-Wallis H = 12.80, p = 0.0017, ε²H = 0.020), accounting for 1.4 to 2.0% of total variance. Tukey HSD pairwise testing localizes the signal to a single contrast between clean limestone and mixed carbonate (p-adj = 0.016); mixed and tight shaly carbonate classes are statistically indistinguishable (p-adj = 0.482). Well identity, in contrast, accounts for approximately 27% of total Fpor variance (Kruskal-Wallis ε²H = 0.271), an effect more than thirteen times larger than the facies effect. Ward hierarchical clustering on standardized per-well statistics partitions the wells into three internally consistent clusters that cross-cut lithofacies composition. Depth is excluded as a confounding variable through within-class correlation analysis. The dominant control on fracture porosity in this dataset operates at the well scale, not the facies scale. Geological similarity between the target and training wells is the binding constraint on the quality of blind-well fracture prediction. These results indicate that carbonate fracture-prediction workflows should replace facies-only conditioning with well-partitioned validation and well-scale distributional descriptors, since the current three-class lithofacies scheme resolves less than one-fifteenth of the variance that well identity alone explains. Structural compartmentalization is a plausible candidate mechanism, though untestable with the present dataset.]]>
