A second LLM judge failed to corroborate 94% of 140,000 flagged failures from 2.5 million evaluations, indicating false positives dominated the pre-calibration suite¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Re-examining every failure from 2.5 million evaluations over a six-month window with a second LLM judge showed 94% of the 140,000 failures were not corroborated. → Chris Rohlfs 2026

Evidence¶

Chris Rohlfs 2026¶

Chris Rohlfs, Rodrigo Vergara Bosse, Priscilla Hopper, Keanon O'Keefe, Patrick J. Russell. (2026). When Evaluators Cry Wolf: Lessons from Production LLM-as-Judge Evaluation in Educational AI. https://arxiv.org/abs/2609.28478 doi:10.48550/arxiv.2609.28478

q2 · i? · design · r2

Production log analysis over a six-month window in which a second LLM judge re-examined all 140,000 failures from 2.5 million evaluations; the article reports "In 94% of the 140,000 failures, the second model failed to corroborate the issue."

"We applied a second LLM judge to re-examine every failure from 2.5 million evaluations over a six-month window. In 94% of the 140,000 failures, the second model failed to corroborate the issue."

Discussion¶

Related Claims¶

• Calibrating a production LLM-as-judge suite raised per-flag precision from 0.6% to 48.9% while cutting confirmed false positives by 99% — related
• Post-calibration evaluators captured 100% of synthetic egregious failures across all 21 deployed evaluators, up from a pre-calibration range of 73.3–100% — related