Joint questionnaire-based grading of insomnia and anxiety severity is practical only if few items are needed. We developed OrdiFuse, a cross-task ordinal stacking model, together with a eakage-controlled, model-aware item-reduction protocol, in 10,322 respondents with physician-assigned grades. Items were ranked by OrdiFuse’s multitask permutation importance, evaluated on folds excluded from the ranking, and reduced under a rule fixed before the locked test set (n = 2,096) was read. The rule retained 24 of 51 items. On the held-out test set, OrdiFuse achieved a multitask macro-F1 of 0.936 and 95.3% exact agreement, significantly outperforming equally tuned single-learner and stacking baselines (∆ = +0.012–0.017; Holm p ≤ 0.004), although its advantage over grid-selected stacking was not significant (∆ = +0.008; p = 0.17). Across three end-to-end re-derivations, all 24 items were consistently re-selected, and OrdiFuse outperformed stacking and histogram gradient boosting. Item selection remained stable across sub-cohorts. However, cross-task fusion showed no measurable benefit, and internal-external validation revealed performance reductions of 0.04–0.09, indicating limited cross-cohort transportability and no consisten comparative advantage. Insomnia grades were near-deterministic functions of the items, so incorporation bias cannot be excluded. The 24 items form a candidate predictive subset, not a validated instrument; independent external validation is required.