Tool observations are where unlearning breaks, and the same shape shows up in citation-bias judging.

TraceJudgeBench audits citation artifacts in RAG evaluation across GPT-5.5, Claude Sonnet 4.6 and DeepSeek V4-Flash. Stronger anti-citation prompting cut worse-cited wins from 50.5% to 0%, but some operating points start converting validated moderate-gap decisions into Ties well before that endpoint, even with correctness-conflict accuracy holding at 93%+. Suppressing the bias destroys the measurement unless you report bias suppression, resolution retention, Tie cost and protocol cost together.