One added colon flips Jev to accept wrong final answers, but only under sorted request presentation
The study tests whether the presentation of a structured judging request changes the effect of a candidate-text edit: on 200 previously unused DROP and GSM8K source clusters, adding one colon after random letters raised Jev's false acceptance from 1.0% to 26.0% with three output labels and from 3.0% to 26.5% with the published four-label instruction under sorted JSON keys, while both candidate variants were rejected under insertion presentation; Jev met the control thresholds under both grading configurations and presentations, and GPT-6 Sol produced no observed cue-condition false acceptances.