Score the candidate against the 10% and two-point rules

Score the candidate against the 10% and two-point rules

Compare the candidate system against the pinned baseline on the untouched test split. Require at least a 10% relative word error rate reduction overall and no dialect group worse by more than two percentage points. Deliverable: A scorer that prints both rules as pass or fail per dialect group and rejects any run whose normalization or split differs from task 1.