When a single adaptive engine serves 22 apps across four certification domains, which domains concentrate misconceptions on hard items, and can this be predicted without learner data? Across 35, 250 production items, score-level differentiation (4. 5× misconception ratio, d ≥ 1. 99) is domain-invariant by design, but item-level concentration depends on a two-factor interaction between difficulty spread (σ) and base accuracy (bd). When domains share bd = 0. 72, wider σ drives 69% more hard-item misconceptions; when σ is nearly matched, a one-point bd difference shifts the ordering. A predictor P (dᵢ > 0. 6) × (1 − bd) ranks all four domains correctly across a 6× sensitivity sweep, enabling domain-aware scheduling from item bank statistics alone. Submitted to EDM 2026 Poster/Demo Track.
Anthony Perry (Sat,) studied this question.