Replication package for evaluating local language models as triage judges for real vulnerability policy evidence

This archive contains the replication package for a real-advisory study of local language models as bounded triage judges for policy-oracle-relevant vulnerability evidence. It includes an 800-case public advisory benchmark derived from NVD, GitHub Security Advisories/OSV, CISA KEV, and CWE mappings; prompts; local model outputs; deterministic, supervised, and hybrid baselines; a masked 100-case label audit with three local judges; parsing and analysis code; derived tables and figures; reproducibility instructions; and integrity hashes. The archive excludes manuscript source, compiled manuscript artifacts, journal templates, submission notes, cover letters, response letters, legacy datasets, and local build artifacts.