Analytical synthesis of these 2026 sources: BioMysteryBench’s historical panel result concerns finding answers from scratch, not explaining a supplied answer’s scope and assumptions. Its v11 audit checks the evaluator: wrong keys or missing information can mimic solver failure, while scoring clarification concerns what earns credit. Brodeur’s 2024 experiment separates two validation targets: recovering numbers and detecting errors that could change claims. High recovery in human-only and assisted teams did not imply equally effective major-error checking. None of these outcomes directly measures explaining a supplied AI result. The 30% attempt-success rate, nine task removals and errors per team cannot form one oversight-failure rate. Whether reviewers can explain and validate supplied AI results remains open here.