SAME WORK, TWO JUDGES Case / version: Real redacted case or invented example: Task and pass criteria: Expected verdict + evidence, or UNRESOLVED: Exact evidence / derived signals provided to both: Judge A model / configuration (or unknown): Judge B model / configuration (or unknown): Judge A verdict / score / rationale: Judge B verdict / score / rationale: Missing responses / errors: What checking the source establishes: Reason for disagreement (or unknown): Specific next change: Fresh case reserved before tuning: BATCH TOTALS Cases attempted: Cases with both verdicts: Agreements / cases with both verdicts: Judge A known-bad cases with a returned verdict: Judge A false passes / known-bad cases with a returned verdict: Judge A known-good cases with a returned verdict: Judge A false fails / known-good cases with a returned verdict: Judge A missing responses / errors: Judge B known-bad cases with a returned verdict: Judge B false passes / known-bad cases with a returned verdict: Judge B known-good cases with a returned verdict: Judge B false fails / known-good cases with a returned verdict: Judge B missing responses / errors: Unresolved expected labels: Missing responses and errors: What remains untested: Guide: https://www.nonstopinfopublishing.com/guides/compare-ai-judges/