Only OpenAI has committed to the strongest approach: testing task-specific versions of its models. But evidence of their follow-through is limited.