AI Role Eval Scores Voided After Export Left 118 of 192 Answers Empty

A company that ran a full evaluation of eight AI roles, covering 192 questions, saw scores ranging from 70.8 to 97.2 before an independent review found the problem. When the exported results were re-read, 118 of the 192 answer fields were empty, and 85 of those empty answers had nonetheless been awarded full marks. The failure occurred in the export step that moves results to persistent storage, even though the questions, model calls, judging, and aggregation all reported success. The company withdrew the entire score table as an official baseline the same morning, recording the correction in the same ledger as the original claim. The author argues that evidence which cannot be re-read is not valid evidence, and that the scarce discipline in evaluation is preserving a traceable chain of custody after generation.