Reproducibility Evidence for Domain-Disjoint Evaluation of the PhiUSIIL Phishing URL Benchmark

Frozen reproducibility evidence supporting the InCoWoCo 2026 conference manuscript ‘When Random Splits Mislead: Domain-Disjoint Evaluation of the PhiUSIIL Phishing URL Benchmark.’ The deposit is not a new release of the original PhiUSIIL dataset. It contains the frozen primary P1–P4 evaluation evidence plus separate authoritative Phase 7 feature-forensics and Phase 8 domain-disjoint feature-ablation archives. The study covers conservative canonical-URL deduplication; record-random and registered-domain-disjoint protocols; fixed-seed evaluation; prediction-derived metrics; bootstrap uncertainty and paired protocol comparisons; seen-versus-unseen-domain diagnostics; lexical sensitivity; feature forensics; and targeted ablation. The evidence demonstrates extreme internal PhiUSIIL benchmark separability under the evaluated controls. It does not establish target leakage, causality, memorization, or external, temporal, zero-day, campaign-level, or deployment generalization.