Data and code for "Benchmark Reuse and Endpoint Drift in LLM-Based Vulnerability Detection: An Empirical Study of Four Chinese APIs"

Dataset and analysis code for the manuscript of the same title. The study evaluates four Chinese LLM APIs on Java vulnerability detection with the JavaVulBench benchmark, in two collection rounds: 3,200 calls on 2026-07-27 (four APIs, three prompt strategies, two data splits) and 4,600 calls between 2026-10-03 and 2026-10-06 (grid re-run, same-period direct-verdict panel, a fifth vendor, a successor endpoint, a third snapshot, and an 800-call GPT-4o replication). The record contains the archived raw responses and the corrected per-record verdicts, the official 200-sample probes, the fine-tuned encoder predictions, the training/evaluation overlap audit, and the code that regenerates every table and figure: the verdict parser with its acceptance test, the metric and bootstrap routines, the table and figure generators, the encoder training script, the non-LLM baseline, and code/reproduce.sh, which runs the CPU-side pipeline end to end. Table 5 of the manuscript is reproduced with three rows; the kimi-k3 run, collected at the provider-mandated temperature 1, is reproduced separately and reported in the manuscript's Appendix C.