Educational decision support requires more than an accurate model: it must separate predictive usefulness from causal interpretation, quantify uncertainty, and connect model evidence to defensible actions. This study evaluates repeated out-of-the-ordinary prediction of self-reported perceived employability and develops a bounded, uncertainty-aware analytics framework for vocational education. The analysis used the version-4 public workbook containing 266 Grade XII students from 17 schools in seven districts of Aceh, Indonesia. Mean perceived employability was predicted from mean internship experience, 21stcentury competence, and psychological capital. Data integrity, response quality, reliability, and construct separation were audited. A mean-only baseline, ordinary linear regression, Elastic Net, and Random Forest were compared using five-fold cross-validation repeated 20 times; tunable models used nested four-fold inner validation. Linear regression achieved MAE = 0.264, RMSE = 0.350, and R^2= 0.643, reducing MAE by 38.8% relative to the baseline. Elastic Net was effectively equivalent, whereas Random Forest had 9.6% higher MAE. Repeated outoffold dropcolumn analysis showed that competence supplied the clearest unique information: removing it increased MAE by 0.0366 (13.8%; paired-bootstrap 95% CI [0.0162, 0.0562]). Psychological capital provided smaller supportive evidence, while internship experience had a limited unique contribution. Bootstrap item-ranking stability identified project execution and leadership as the strongest curriculum-review priorities. The contribution is an auditable data-to-decision protocol that integrates nested predictive validation, conditional construct evidence, measurement diagnostics, ranking uncertainty, calibration audits, and explicit decision boundaries. The framework supports curriculum and program review, not causal prescriptions, automated placement, or high-stakes individual decisions.