Aggregate refusal scores on mixed safe and unsafe prompts can obscure local boundary failures and label-correlated metadata shortcuts. We present PAIRSAFE, Proxy-Resistant Paired Safety Evaluation, a score-agnostic diagnostic protocol built from 150 deterministic, matched XSTest-derived safe/unsafe pairs. PAIRSAFE distinguishes strict pair locks from audited residual fields, then asks whether nuisance proxies collapse, unsafe twins rank above safe twins, and one threshold yields pair-consistent decisions; generated behavior is evaluated separately. Strict source, normalized-type, and pair-ID controls collapse to PSA 0.50 and PCDA 0.00, as required by construction. Residual metadata remain informative: a leakage-safe combined probe obtains nested Cal-PCDA 0.397, yet residualizing model scores against measured metadata leaves substantial paired signal. Direct-logit margins for Qwen2.5-3B-Instruct and SmolLM2-1.7B-Instruct reach PSA/Cal-PCDA of 0.987/0.880 and 0.973/0.760, respectively. Qwen’s secondary generation check is lower again (PCGA 0.700). These disagreements show why local ranking, shared-threshold decisions, and generated responses should not be collapsed into one score. PAIRSAFE is a bounded diagnostic protocol, not a new classifier or deployment certificate.