Core.today reports that rescoring the same answers in its LoCoMo memory-benchmark experiment produced scores from 55.8 to 81.4 depending on the judge prompt. The analysis points to answer-key errors and ambiguous judgments as sources of variation alongside model choice. A separate public audit repository records 99 answer errors among 1,540 questions plus additional citation problems. The audits use different conditions and counting methods, and their results were not reproduced here. System comparisons need to disclose judges, prompts, answer-key revision history and costs together.