A refusal score can rank two prompts correctly and still misclassify the safe one. PAIRSAFE follows both decisions across 150 matched pairs, comparing scores, held-out thresholds and generated responses.
The generation check covers Qwen and uses archived rule-based labels. The release reproduces the main table from scores and identifies the final nested metadata audit as a transcription from the paper.
Qwen ranked 148 of 150 matched pairs correctly. A threshold fitted on other pairs classified both members correctly in 132. PAIRSAFE traces the difference from individual scores to threshold decisions and generated responses.