Evaluating an answer you did not write
A short checklist for reading a cited regulatory answer under time pressure, whether it came from a colleague, a consultant or a research agent.
You will spend more of your career reading other people's research than writing your own. The skill is not 'does this sound like a person who knows the area'. The skill is 'can I reconstruct this in ten minutes, and if I cannot, what is missing'.
Five checks, in order
- Question fidelity. Does the answer address the question that was asked, including the jurisdiction and the product form, or a nearby question that was easier.
- Cite quality. Instrument, version, internal address. A document title is not a cite. A case name is not a cite.
- Overreach. Does any sentence go past the paragraph it points at. Conditions, cut-offs and notes are where overreach lives.
- Silence. Are gaps labelled as gaps, or filled with 'generally' and 'typically'.
- Trail. Can you see what was searched and what was opened, including empty results.
If any of the five fails, the answer is not ready to leave the building. Fluency does not rescue it. A well-written wrong location is worse than a clumsy right one, because it travels further before someone checks.
Do not outsource the click
Open one cite. Not all of them, one. If that one does not land on the claimed sentence, stop trusting the rest until you have opened them too. This is slower than reading the prose, and it is the only check that catches invented or drifted sources.
If the product will not let you open the cite from the page, treat that as a failed check. A bibliography you cannot follow is decoration.
What good looks like
A short answer, a small number of cites that actually carry the claims, a trail a second person can walk, and an explicit statement where the record does not speak. RegAffairs AI is built to produce that shape. Demand it from any tool, colleague or firm that wants you to quote them.