The UK's nuclear regulator published a report earlier this year that names the governance gap more precisely than anything I have seen from the AI policy community. And then provides no instrument to close it.
The Office for Nuclear Regulation's sandboxing report sets out a three-claim assurance framework for AI in nuclear installations. The first claim an operator must demonstrate is that "the requirements are sufficiently defined for the system." The second is that the system meets those requirements at deployment. The third is that it continues to meet them over time.
That first claim is where everything starts. If the requirements are not sufficiently defined, claims two and three cannot be demonstrated regardless of how well the system performs. The ONR has named the foundational obligation correctly.
It has not said how to check whether it is met.
There is no instrument in the report, and none elsewhere in the AI governance literature, for assessing whether a requirements document is structurally complete before it is used as an assurance basis. The obligation exists. The method for discharging it does not.
This is not a failure of the ONR report. It is a precise description of the gap. The most safety-conscious regulatory body in any jurisdiction has confirmed, in its own language, that the governance document quality problem is real, structural, and currently unaddressed.
PromptQ is built to discharge that first claim. The waitlist is at promptq.ai.