Auditing AI C++ Patch Graders by Seeding Faulty Mutants

A hypothetical example describes an eval harness that reported PASS on AI-generated C++ patches for weeks while a signed modulo replacing a bounded comparison went undetected. The unit tests and sanitizer checks stayed green because no test case exercised the faulty branch, showing that a passing report does not prove the grader can still catch a given failure. The proposed remedy is a weekly self-audit that pushes known-faulty patches, called mutants, through the real harness and flags any grader that fails to reject them. Each mutant is declared in a manifest with the graders that must catch it, and a surviving mutant counts as a finding. The approach requires a clean git worktree and a harness that reports per-grader status in machine-readable form.