The original title is: "Six experiments on adversarial verification — and the 75% wall that didn't move"

Short summary

Six experiments reveal that LLM-based adversarial verification hits a systematic 75% false-negative wall that no standard technique — rerun voting, multi-prompt voting, or prompt calibration — can move. The wall exists because a reviewer draws a boundary line: sharpening it catches more garbage but rejects more valid output. The practical implication is to stop trying to eliminate the line and instead design around it.

• •LLM reviewers hit a stable 75% false-negative rate — rejecting 3 of 4 valid outputs — that voting and prompt calibration cannot shift
• •The wall is geometric: sharpening the review line reduces false positives but increases false negatives on the same curve
• •Smaller models aren't better verifiers; they sit at a different operating point on the same precision-recall tradeoff

Generated with AI, which can make mistakes.