Replication package for "Detection Rate Is the Wrong Metric: Adversarial Comment Injection Shifts the Decision Threshold of LLM Code Reviewers": verdict logs, scripts and checks

Replication package for the paper "Detection Rate Is the Wrong Metric: Adversarial Comment Injection Shifts the Decision Threshold of LLM Code Reviewers" (accepted at ICAISG 2026). Contents: raw verdict logs for a 2x2 evaluation (vulnerable and safe C functions, clean and under four comment injection attacks) on five sizes of Qwen2.5-Coder (0.5B to 14B), a chain-of-thought probe, and a replication with #, // and /* */ comment forms on the 7B and 14B models; the scripts that produced them; and verification scripts that recompute every table in the paper from the logs. Each call stores one parsed verdict. The Devign function text is not redistributed; the loaders rebuild the function sets from the public Devign dataset. Known limits (an unused incomplete deepseek-coder:1.3b run, 20 missing replication calls, no raw responses stored) are listed in README.md inside the archive. Data and documentation: CC BY 4.0. Code: MIT.