Early detection of type 2 diabetes (T2D) is a public-health priority, yet externally validated, blood-based screening tools remain limited. This work implements a fully reproducible biomarker-discovery pipeline using whole-blood microarray data, trained on GSE9006 (Affymetrix GPL96; Ntrain=117, Control/T2D=105/12) and externally validated on GSE15932 (Affymetrix GPL570; Nval=16, Control/T2D=8/8). The objective was to derive a small, interpretable gene panel for T2D classification and to quantify performance with rigorous uncertainty estimates. Probe-level intensities were mapped to HGNC symbols and collapsed to the gene level. Differential expression (limma) was performed within cross-validation folds to avoid information leakage. Model-agnostic stability was quantified via Boruta and bootstrapped elastic-net (B=500), and the combined evidence yielded a 12-gene candidate panel (e.g., ETS1, GJD2, RALGDS). To mitigate cross-study heterogeneity, the combined train/validation matrix (restricted to the panel intersection) was adjusted with removeBatchEffect; PCA before/after adjustment showed attenuation of train-validation separation while retaining class structure. A shrunken elastic-net classifier was refit on the panel and evaluated in an independent cohort. Performance estimates were as follows: nested cross-validated AUC in training 0.877; external AUC 0.562 (95% CI 0.235-0.890; bootstrap B=2000). Using the Youden-optimal threshold, sensitivity 0.375 and specificity 0.250 were obtained; assuming 11% screening prevalence, PPV 0.236 and NPV 0.942 were achieved. Calibration was modest (Brier score 0.629), and a label-permutation test indicated no evidence above chance (p=0.677). These findings indicate modest discrimination in the independent cohort, consistent with platform and cohort heterogeneity, while delivering a leakage-aware workflow, a stabilized small panel, and clinically interpretable metrics that can guide multi-cohort refinement and prospective assay development. The full code and intermediate artefacts are provided to facilitate reuse and extension.