Claim: reliable improvement. Evidence: 82/100 versus 80/100 on one set. Concern: uncertainty and run variability are not reported; reliability is not established. Request: paired predictions, test construction details and repeated-run or suitable paired uncertainty analysis.
