Expert attention is scarce, yet clinical artificial intelligence often escalates by fixed thresholds. Hereby present Calibrated Refresh Systems (CRS), a systems-level design pattern for calibrated expert review rather than a new estimator or validated intervention. CRS accepts a low-resolution action only when calibrated residual bounds cannot cross an action threshold; otherwise, the case receives high-resolution review. It couples selective prediction, learning to defer, conformal calibration, reference-label governance, accepted-case audit, drift arbitration, subgroup safeguards, and never-fade rules. A synthetic capacity-matched mechanism check illustrates how harm-aware budgets can reduce a prespecified weighted disagreement proxy without implying improved patient outcomes. CRS yields testable operational artifacts and a staged validation roadmap, generating artifacts suitable for safety governance and regulatory documentation while preserving attention for rare, high-harm cases. Its claims depend on exchangeability, adequate calibration samples, credible reference processes, workflow capacity, and prospective clinical evaluation.